📝｜prompting-help

[sick cipher]

Oh yeah

Is that the only way ?

[wheat swift]

You can put a png with the same name as the the ti in your embeddings folder I think

[sick cipher]

That's smart

[wheat swift]

Probably easier to generate an image. that way it's generated from the TI

[sick cipher]

Yep

[wheat swift]

Trying to figure out what I want to make. Already did a poster today

[sick cipher]

@wheat swift do something with a bat

[wheat swift]

another one?

[sick cipher]

how many did you make ?

Alright here is something I never seen anyone do

Do a Uganda knuckles but a good one

Like the live action sonic style

Oh just saw the bat pic looks wierd af

was that your attempt of a knuckles? @wheat swift

[wheat swift]

That was SDs attempt. Not my normal subject

[last grove]

#1072220168534642768

[sick cipher]

any tips to get rid of the seem like line between the original and the outpainted pixels ?

[wheat swift]

yeah. just inpaint over them

make sure you turn off outpaint first though. been down that road a time or two

[sick cipher]

wait what ?

[wheat swift]

turn off outpaint before you inpaint or it will outpaint again

[sick cipher]

which part do i inpaint

[wheat swift]

inpaint over the seam. go a bit wide and it will smooth out the seam

[sick cipher]

With the outpaint is it fine for the denoise to be 1 ?

[wheat swift]

I've never set it that high. Usually about 0.75 is fine, is it working for you?

[sick cipher]

it was 0.8 now trying 1

Nope a disastrous

[wheat swift]

would be better to set the mask smoothing larger then the denosiing

[sick cipher]

Is the mk2 better then poor man's outpainting ?

Or each has a use case ?

[wheat swift]

I prefer mk2 over poor man's

did your outpainting work except for the seams? you have enough vram?

I'm going to talk you through using img2img as glue if both those are ture

[sick cipher]

I only have 4GB Vram 💀

now we have two seems lets goo

[wheat swift]

how big is your picture? you have to do a small batch, but this will work

[sick cipher]

512x512 batch count 1 batch size 1

[wheat swift]

send the picture to img2img. don't change your prompt, set your denoise to 0.25, click generate

the low noise makes sure it doesn't change the picture too much, but it will bind it all together and make it fit

[sick cipher]

How wmuch sampling?

[wheat swift]

sample steps should stay the same. everything else the same as the original generation

except the resolution. that will change to match the picture you send

[sick cipher]

Seem is gone but now I have a somewhat different picture so I will rise the denoise a bit

[wheat swift]

no you want to go down on denoising to keep the picture from changing

[sick cipher]

Thank you for everything I have to go now I will miss with more later

[sick cipher]

what do I do then ?

inpaint ?

[wheat swift]

inpaint is a sure fire way to make sure it is the same and only fixes what needs fixed

it's just a lot more work. It's my normal workflow

[sick cipher]

What if I redo the outpaint the original pic but with the seed will that help ?

[wheat swift]

You may still get a seam. It's just the nature of the beast

[sick cipher]

Aight then see you later

[wheat swift]

Take care

[scenic walrus]

Hello guys, do you guys use stable diffusion too?
I'm quite new to it and trying to make clothing damage image.

I was wondering if anyone know of a way you can manipulate the prompts to have more control on how a person's outfit is damaged, to what degree, in what way or if possible "what's damaged, and what's not"?

Anyway, thanks in advance.

[runic osprey]

have you tried adding parenthesis around damaged? i think it makes that term stronger. i think it also depends how far it is from the start of the prompt, so if it's at the end it wont have as much of an effect

[low ivy]

hey there, i am looking for promts to generade a realistic 9ft tournament pool table, some ideas ? i have struggle with dimensions, to much pockets...

[dawn lagoon]

Have you tried making an image "clothing", then using the inpaint tool to draw over the areas and run it at full resolution with the prompt similar to "damaged clothing" or "frayed clothing" until the areas you want damaged look good to you?

[past sluice]

Hopefully this isn't too redundant a question, but for the life of me I can't make 2.1 generate a "hypodermic syringe" for a medical job, is this somehow disabled within the model itself?
If so, is there a list of items documented somewhere that are explicitly disabled like this that I can provide as a reference for the reason the job was not possible to complete?

[scenic walrus]

Hello skeddles,

thanks for the tip. I actually did try this before. This does have it's effect. In fact, you can go from (ripped) all the way up to (ripped:1.5). Anything above 1.5 makes the image look weird.

However, at the very beginning, I was having a hard time even getting damage on the clothes at all. Despite trying the () modifiers. After a lot of trail and error, I realized, the program was not reading terms such as ripped(they read this word as muscular) and torn. Turns out, something that makes a huge difference is the symbol underscore _

When you say ripped_clothes. This makes the program understand and adding (word:multiplier) does the trick. Which is where I am currently at.

However, the issue I face currently is.

1. You know how increasing the factor increases the subject amount. For example, if you put (((((dog))))), there will be a lot of dogs in the image. So, for my images of torn clothes, the clothes will be ripped. But SD would add a bunch of extra fabric to the clothes. And a simple shirt, for example, would not look like a shirt. It would look like a shirt with a bunch of extra fabric attached to it but is badly torn.

2. I am unable to control how the damage goes. For example, If I want a missing fabric from someone's right shoulder, entire right half of the shirt ripped off(like Goku tends to be like), or something very specific like the character just wears 2 shirts, the one underneath is fine, and the 2nd has only scraps left.

There would be very little control on what I can do, SD seems to affect either all or nothing.

3. Lastly, the tear and damage always looks unnatural, but this is a nitpick as SD already does so much for you.(I'll try to fix this as the last thing)

[scenic walrus]

coincidentally, I just happened to encounter a reddit thread that tackled these functions.
I did play around with image to image yesterday. I saved an image of Goku with his right side of the shirt torn off and did my usual prompt. Unfortunately, it changed the whole damage to the usual weird looking/uncontrollable tear my promps usually appear as.

However, what you said, is actually a good idea I should try. I actually didn't think of doing this at all. Using the inpaint tool in certain regions. Huge thanks for the suggestion, really. I will try it out and let you know how it goes.

[dawn lagoon]

Another tool you can use is ControlNet, which allows you to use a reference image for the purpose of composition. You can use it in conjunction with img2img or even txt2img. I haven't tried using it for details like ripped clothing but you may find it possible to replicate the rip patterns of images online and put them into your desired image.

[earnest pawn]

Any tips for generating interior design photos? I want to try having Stable Diffusion make my dream house.

[atomic flume]

Hey I'm having a problem with a prompt, I'm currently trying to generate a single eliete female sci fi solider in an image but when I do it two always pop up

(extremely detailed CG unity 8k wallpaper), (one:1.3) (elite female sci fi solider:1.2) in sleek form fitting intricate power armor, sci fi assault rifle, professional majestic oil painting by Ed Blinkey, Atey Ghailan, Studio Ghibli, by Jeremy Mann, Greg Manchess, Antonio Moro, trending on ArtStation, trending on CGSociety, Intricate, High Detail, Sharp focus, dramatic, by midjourney and greg rutkowski, realism, beautiful and detailed lighting, shadows, by Jeremy Lipking, by Antonio J. Manzanedo, by Frederic Remington, by HW Hansen, by Charles Marion Russell, by William Herbert Dunton, (No Helmet:1.4), (sci fi city:1.2), walking in a warzone

Negative prompt: western, cowboys, hat, disfigured, kitsch, ugly, oversaturated, grain, low-res, Deformed, blurry, bad anatomy, disfigured, poorly drawn face, mutation, mutated, extra limb, ugly, poorly drawn hands, missing limb, blurry, floating limbs, disconnected limbs, malformed hands, blur, out of focus, long neck, long body, ugly, disgusting, poorly drawn, childish, mutilated, , mangled, old, surreal, text, (multiple people:1.2), multiple subjects
Steps: 60, Sampler: DPM++ 2M Karras, CFG scale: 7, Seed: 1161103013, Size: 1080x1272, Model hash: bc561295ca, Model: protogenInfinity_protogenX86

[earnest pawn]

I usually have that problem when I change the dimensions. The AI was trained on 512x512 images so anything bigger than that and it apparently just starts tiling?

[atomic flume]

Oh

ok

[smoky totem]

try inpainting to localize the area where ripping hapens, you may just "paint" the ripping trail just the way you want it with a finer brush size

you may have to tweak the prompt to just say (ripped_clothing) and get rid of the character and background description etc.

I always use a DOC to write the prompts, keep them in version numbers and then the images I keep always have a version number trailing them so i know what prompts I used to generate those in the future, in case I want to re-roll or make alternatives of those. I should probably make an excel file instead of a DOC.

I know there is a way to extract prompt info from an image, but i am not sure if they are good at extracting all the inpainting, outpainting, re-rolls that an artist may have prompted in order to create the desired image. Many prompts I see from galleries like the ones on playgroundai simply show the final prompt, which sometimes look so empty that it seems to come out of an inpainting process, like ripped_shirt colar. Whilst the picture show an epic battle happening lol

[proud gust]

Hi! Can someone help me with the correct syntax for blending between two things, like if i wanted to mix a cat and a dog by percentage, or two different embedded face trainings together, or an embedded face and a lora - whats the terminology to do this precisely

[wheat swift]

[object1:object2:0.5]

it's called a prompt edit

There are other ways to do it too. hold on. Getting the docs

https://github.com/AUTOMATIC1111/stable-diffusion-webui/wiki/Features#prompt-editing

https://github.com/AUTOMATIC1111/stable-diffusion-webui/wiki/Features#alternating-words

and here is the other one I was looking for https://github.com/AUTOMATIC1111/stable-diffusion-webui/wiki/Features#composable-diffusion

[proud gust]

amazing, thank you, i will dive in and research this now

Second question; is it possible to restrict certain attributes to one object or character, so this prompt doesn't bleed in to other thing, like say you wanted someone on a green chair, occasionally it does stuff like make their hair green, or give them a green shirt or something, is there a way to ring fence attributes to only a specific set of defined objectds

[wheat swift]

Only way I have gotten that to work is work on a section at a time

[proud gust]

obviously there's inpainting but frustrating to control an image, like if you wanted lets say a normal looking person in a blade runner style future city its hard to not get it to put cyber crap on the character

(actually not tested that example but you know what i mean)

[wheat swift]

In painting is the only way I have gotten that to work

[proud gust]

even if you say "there are photos of X on the wall in the BG" it may still put X on the characters T shirt or on the desk they are at or whatever

[wheat swift]

sd in notoriously bad at color control

[proud gust]

AssertionError: AND is not supported for InstructPix2Pix checkpoint (unless using Image CFG scale = 1.0)

interesting

[sharp aspen]

hi, is there anyway to colour a bw image using img2img?

[proud gust]

is there a command to prevent instructpix2pix from changing a specific element in an image, like "do not change the shirt" or "preserve the color of the hat" etc

[wheat swift]

You could probably do some creative prompting, but deoldify is a neural network trained to do just that

[proud gust]

instead of putting text that attempts to say "do not change shirt" on the shirt lolllll

[runic osprey]

text is a no no

[vapid moat]

maybe that trick I have seen people doing with having an img2img input as well as the controlnet input, if you set it the same it should theoretically keep some of the stuff

[copper portal]

Looking for a prompting or similar method where I can specify goals and constraints, such as performance, materials, and manufacturing. It could be useful for automotive, aerospace, defense industries...

[craggy steppe]

How can I get this kind of result in Stable Diffusion? Similar style but long range photo. Tried few prompts but it just changes the cloths into random colors

[ionic wing]

test

[atomic flume]

Hey I'm trying to generate norman rockwell style stuff

but when I do it I keep getting weird extra limbs and stuff any suggestions here are my prompts

norman rockwell style:1.2) illustration, excessivism, (a beautiful woman:1.2) in a sundress while she drinks a bottle of soda on a Saturday afternoon, golden hour

Negative prompt: Deformed, blurry, bad anatomy, disfigured, extra limb, ugly, poorly drawn hands, missing limb, blurry, floating limbs, (mutated hands and fingers:1.4), ((anthro)), ((animal)), crown, flowers, candle, fire, flame, hat, horse, riding, umbrella, snow, zipper, zip, logo, text, water mark, (to many fingers:1.3), (bad hands:1.4), (disconnected limbs:1.5), weird hands, (multiple legs:1.2), floating limbs

Steps: 60, Sampler: DDIM, CFG scale: 7, Seed: 704020129, Size: 616x816, Model: protogenInfinity_protogenX86, Denoising strength: 0.7, Hires upscale: 2, Hires upscaler: Latent

[steep aurora]

Hey! I'm looking through the documentation trying to find the difference between putting a comma ',' and an AND in a prompt. For example, how does stable diffusion treat "cat, brown" different from "cat AND brown" ?

I know AND is for 'composable diffusion', but I'm trying to figure out how it's useful.

This is what I was able to find, but I'm looking for more of an explanation.

is the difference that it diffuses each prompt separately? So it adds each prompt to the previous result? Is there a way to apply a set of prompts to all AND statements?

Such as:

(cat:0.3 AND brown:0.7), sharp focus, illustration, etc?

[pine spindle]

I might be wrong but my impression is that stable diffusion does not particularly consider commas.

Even withouth commas you get something very similar.

[lavish surge]

Hello there, I got a question,

How do you make the ai able to identify which anime character you are going to make it to generate?

for example for those characters who have a name only

like Reze from chainsaw man

[runic osprey]

it probably wont be able to create a specific character for you, especially an obscure one. I'd either describe the character to the ai, or do some finetuning

[lavish surge]

alright, thanks

[pine spindle]

You can download a model that has the character.

https://civitai.com/models/6529/reze-chainsaw-man

[lavish surge]

oh, almost forget something like that exists lel

[tired vigil]

I am trying this prompt : 3d render of batman in a shiny black armour with a bat logo on the armour which is glowing blue... unable to get a proper generated image ... all give a batman with a bat logo but it is not glowing and it is not blue ... any suggestions ??

[runic osprey]

did you try emphasizing like (glowing blue)

or try rewording it, like with a glowing blue bat logo

[tired vigil]

Thanks a lot for the suggestions, but they did not turn out as expected , also tried : 3d render of batman in a shiny black armour with a (glowing blue) bat logo

[runic osprey]

its often hard to get it to add details exactly where you want them. inpainting might help show it where you want it, or photoshopping a blue glow on and doing img2img. i havent done much of that though.

[solar violet]

any one have a ghibli prompt i could use

[lavish surge]

Can lora models compatible well with the major realistic models?

[wheat swift]

if it is trained against a 1.5 model, it will be compatible with a 1.5 realistic model, but the results may not be exactly the same. you may need to increase the weight a bit

[proud gust]

if i wanted to do a person in the spaceship from the movie "Alien" how would i do it without turning the person themselves into some stupid "gray" alien, like SD wants to

like whats the best way to use a specific movie as a reference

[mossy cloak]

Hey folks, I’m working on creating an app for professional prompt engineers to better help them create, share, and track prompts along with their renders. Question for you: How do you keep track of your past prompts, settings, and outputs right now?

[proud gust]

they are baked into the png files so i just load files into PNG info and extract them that way, so if i want to take settings from an image that worked well, load it in that way

[earnest dune]

I’m using image-to-image processing to modify pictures of room interiors. Any ideas which prompts I can use to NOT alter structural elements like windows, doors, beams, etc.? Maybe negative prompts can help?

[vague flax]

get the controlnet extension

[robust tapir]

Yes, like the most popular images are made from koreandoll Lora on top of chillmix model

I trained one with as few as 15 my own photos and then one with 120 photos, both worked perfectly with chillmix. Orange mix models are best for anime

[calm marsh]

Hi, I'm new to AI stuff. Came from trying free Midjourney and now install a local SD.
My problem is that, no matter how I tried I cannot get a good result with SD compared to Midjourney.
I don't think at all it is the SD fault. But just can't get anything good even using the prompt from Lexica.art

[autumn jewel]

What models are you using and what kind of effect are you trying to get? you can also try embeddings prompts and negative prompts are very important as well

[autumn jewel]

Scroll down and you will see models and embeddings tab its like two below the one we are in now shows a bunch of different models and embeddings you can use on the side bar to the left

[calm marsh]

Sorry I might give you wrong info. pretty new to this stuff.
I use Stable Diffusion v1.4 original. I just follow one youtube how to install that locally.

[autumn jewel]

you can use any model with stable diffusion one sec I'll find a video

[calm marsh]

Yes I actually onto that. as I understand the model, if it is .ckpt I could put them on the model directory.

But because I'm new and no coding background, I still don't know how to get other thing to work, like for ex. I have one .pt file and don't know how to use that at the moment.

[autumn jewel]

ok you can put safetensors in the model folder too I hear that's best but I use ckpt as well one sec I'll look up the pt folder I think those are
files

[harsh crane]

Ia this an original stabel diffusion?

[autumn jewel]

yes .pt files are embeddings

[calm marsh]

so I just put .pt files in the same place for .ckpt (model) ?

[autumn jewel]

just put those in the embeddings folder

[autumn jewel]

Nooooo sorry put them in the emnbeddings folder

[calm marsh]

Ah okay, I see it now. So those one need to be trigger, meaning use the trigger word in prompt, yes?

[calm marsh]

https://civitai.com/models/6543/old-fashioned-diffusion it is this one I want to try at the moment.

[autumn jewel]

yes teh models work best witht he trigger word too Nitrosocke has some really good models

[autumn jewel]

Ahhh yea I liked that one too

[calm marsh]

So, basically models will narrow the result down to what you are looking for, and those embedding will help at well? as my understanding.

[tired vigil]

Hi

[autumn jewel]

Yea they both help with getting good effects there are a lot out there

[autumn jewel]

I have to go take a shower and get ready for work I hope you enjoy and remember prompts matter a lot negative prompts as well

[calm marsh]

So just to let you know where I came from, the right pic is from SD1.4 I installed today, and the right is the Midjourney before I ran out of trial.
With SD the whole day I got these cheaper looking art see on the left image.

[calm marsh]

Thank you!

[autumn jewel]

Not bad at all for y9our first images you should ahve seen mine way back lol

[calm marsh]

Alright, thanks a million!

[autumn jewel]

Oh and certain embeddins will only work with the model in that how you say version like embeddings will only work with the 1.5 or 2.1 model so when pairing an embedding with a model make sure they are the same version or at least work together most of the time the creator of the embedding will say what version it works with 1.4, 1.5 2.0, 2.1

[autumn jewel]

wait until you get to the point to where your merging the models its so much fun seeing what effects you can get

[calm marsh]

Yeah I saw that on youtube as well. I guess let's see if I get any decent with these one first.

[autumn jewel]

you will it's not hard just remember prompt is super important

[calm marsh]

Yeah, at one point I ran out of word. I saw stuff like more than just "subject" but some symbols... I won't be taking more of your time atm before you late for work! thanks again!

[blazing quartz]

@calm marsh think of Midjourny as Apple, and SD as Linux. Midjourney gives you everything slick and complexities hidden from you, but you get what they give you. SD gives you all the tools for you to use, including screw yourself up. But you have much more flexibility on things like what models to use, controlnets, etc etc

also, don't forget negative prompts

negatives are just as important as positive prompts

[calm marsh]

Hey, yeah I totally got that right away that's why I wouldnt say it is SD fault that I didn't get good result. But yeah I guess learning curve, but with other models+embedding now I feel I get a better result already.

[blazing quartz]

I would suggest going to some place like civitai, see what models are out there. Most model pages will have images generated with the model, with the positive and negative prompts listed if you mouse over the !, or if you click on the image

that'll help with you getting started with prompting

[calm marsh]

Hi, what is the best negative prompt to avoid the face/head being cropped out? thx.

[empty tundra]

I'm am artist who wants to start using Stable Diffusion as a way to come up with ideas and compositions, but where I can later redraw those generated compositions myself. This also means that the generated images don't need to look good, they just need to have interesting "ideas". I however can't for the life of me get SD to generate interesting images that are still diverse. Prompts like "interesting composition" of course don't work, and whenever you do get interesting results it's when the prompt is so specific that all the generated images look nearly exactly the same.

The new --chaos feature in MidJourney seems really interesting so I might swap over to experiment with MidJourney instead, but I'd love to get some tips on how I could achieve the same with SD

[blazing quartz]

I use "cropped", "out of frame", but not guaranteed

[blazing quartz]

maybe ask ChatGPT to generate you some random prompts

[calm marsh]

Any suggestion of what model to use for the result like the bottom left, a painting style that looks like a retro Gi Joe action figure card? Thanks.

[ruby breach]

hello guys im using the sd and its working nice, the pictures are detailed but low quality, how to get them to proper quality? like ive seen 1080 or 4k these are really bad quality

like this

[runic osprey]

check off hires fix

[ruby breach]

ooo

let me

try

check off means tick it?

[runic osprey]

yes

it should open some additional settings

[smoky totem]

Has anyone found consistent prompt words that tell SD to render a zoomed out view that show a whole subject? With many SD models, I keep running into problems with them showing cropped images. Is it because the models are trained with images of cropped subjects? Here is an example, using cars:

positive prompt: 3dmdt1, raw photo, a futuristic car, luxury, centered, (zoom out:1.8), photorealistic, reflective car body, intricate details, epic, beautiful lighting, best quality, hdr, dtm, (ultra hd:1.1), 100 megapixels, 10mm wide angle lens, view full car

negative prompt: (cropped:1.5), (out of frame:1.5), (zoomed in:1.5), (close up:1.5), washed out, faded, haze, oil, plastic, low res, (worst quality:1.3), (low quality:1.3), stretched, deformed, normal quality, jpeg artifacts, 3d, rendering, drawing, illustration, blurry, crown, hat, black and white, border, frame, lowres, asymmetrical, blurry, disconnected, duplicate, signature, username, frame, logo, (Watermark:1.5), (Text:1.3)

I get mostly images of the 1 picture: a cropped car. and in 1 out of 10-20 images I get a full car but only from front views. If I change the image aspect ratio, ie making the image wider/bigger, I get deformed or stretched cars instead of it showing the full car, despite deformed and stretched mentioned in negative prompt.

Leonardoai the webapp has a button for "zoom out" that fixes this. It is based on stable diffusion 1.5 and 2.1, does anyone know if there are magic keywords that SD immediately understands? because I typed zoom out, view full car, wide angle lens, etc in different orders of the prompt (way up front of the prompt to emphasize, with brackets and numbers), none of those strategies work. It make me suspect if it is due to the model being fed mostly images of cropped subjects.

[runic osprey]

i have the same problem with people. havent really found any good prompts or negative prompts that fix it. making the image rectangular might help a little.

i think it mainly stems from the images SD was trained on being autocropped from rectangular images, meaning they frequently had things chopped off.

outpainting might be able to fix the on the results you like (though i havent tried it).

you might be able to train your own embedding/hypernetwork to fix it. I've been training an embedding based on a certain style and make sure all the inputs were cropped, and I definitely get less cutoff things than before. so maybe you could train one on car images. (though this takes a while, and im not even sure SD would be able to pick up on something like that.)

lastly you might wanna try control nets, which you could use to force the car to be drawn in a specific place (im just not sure how far the design could diverge from the input without affecting the size/position of the subject.

[smoky totem]

yeah thanks for the suggestions. I've trained with other objects with images that I've carefully cropped to show the full object, and some close ups, in different perspectives, top/bottom/elevations and that my custom model is less prone to the constant zoom in view. Unless it is mixed with another subject that maybe trained differently. I will try your suggestions.

I am just curious as to how leonardo fixes it, maybe it is some form of outpainting and then merging the results to give you a "zoomed out" version of your creation.

its just a 1 click operation there.

[runic osprey]

yeah thats what i would guess too

[charred wigeon]

Anyone got some tips for negative prompts I can use when generating a person with a white background ? It keeps rendering the background as a light source, so the edges of the person has a slight white glow. Is there a way to get around this?

[smoky totem]

try white plain background or white canvas/ white studio background?

[charred wigeon]

just on a plain white background, I have an example here

[smoky totem]

put negative prompt backlit backlighting, et

etc

edge glow

[runic osprey]

i didnt have a ton of luck with getting a white background, it's whats i want too. except when I trained my own embedding on images that only had white backgrounds, then i get them more often than not

[charred wigeon]

thanks guys I'll try it out 😄 Some things I render that has a white background turns out just fine, but with this one I got the white edges every time

[proud gust]

is there a good way in SD to have say 4 embeddings and have it randomly use one or the other in the various images it generates? like an "or" command

a list of things this OR this OR this

[blazing quartz]

maybe this will work? https://ericri.medium.com/smile-get-wild-with-stable-diffusion-wildcards-and-facial-expressions-83a52e4ca9fd

I haven't tried it myself, so ymmv

[runic osprey]

dynamic prompts?
edit: just tried it, seems to work. install the extention then just do {embedding1|embedding2|embedding3}

[proud gust]

where do i get this?

[proud gust]

cool will check it out

[runic osprey]

https://github.com/adieyal/sd-dynamic-prompts, just add it in the extensions tab

[blazing quartz]

both wildcards and dynamic prompts are in extensions

[proud gust]

thanks yeah just installed

[proud gust]

has anyone tested Dynamic Thresholding (CFG Scale Fix) and figured out what decent values are to use as starting point

[autumn jewel]

it might be easier to look through model s and find one you want to work with rather than have a style picked out unless your an artist and want to train your own style or you want to collect some images and train a model but maybe https://huggingface.co/nitrosocke mo-di-diffusion or nitro-diffusion and playing in the prompt " retro GI Joe" comic book animated show or something like that might work not sur haven't watched that cartoon since I was a kid

[autumn reef]

has anyone figured out a clear method to getting a consistent full body render? 9/10 of mine are always from the chest up.

[autumn jewel]

https://www.youtube.com/watch?v=xMmCZ1EqMVA Closest thing I have seen

[autumn reef]

thank you. i will look into it

[calm marsh]

Thanks for the tip. yeah, I just like the art style on my uploaded example.
Maybe I rather need to find the right prompt for it as you said. Probably an artist's name or something.
I tried Alex Ross, not exactly the style I search for but maybe close. (example below)
I think his style is still too realistic comic drawing compared to what I want.

[calm marsh]

I went to Nitrosocke as you mentioned, but didn't get to download them yet.
Many interesting ones but models size are big, so I just didn't download everything I want.
Maybe the "classic anim diffusion" would work for what I want. I'll try later.

[tired vigil]

depending on how much source material you have, you could also train a model or a lora specificaly for it too

[mellow hamlet]

thoughts on a prompt to achieve this style?

As in, Im not sure what the style is called or what I would need to include to get a very similar style, so any words that relate to it you might think of could be helpful

[subtle fossil]

hey does abyss orange mix 2 prefer danbooru tag prompts or normal people prompts

[calm marsh]

Thanks, I'm quite new (second day using it) and I heard about train a model.
I understand what is a model and an embedding now. But I haven't look into lora yet.
Will probably get into it eventually. Thanks!

[tired vigil]

well, all trainings work in the same way, you give them a dataset (a series of pictures you want it to learn) and some tokens to learn it on, and they produce a way to then prompt the things you trained on. All work the same on that front. bu all have their up and downsides, depending on what you want to do

LORA is the newest method, quite high on the quality side of things, while being also fast to train and light to share. it's the current craze from what I see.
Full model finetuning can be more qualitative imo, but is longer to train and higher size to share.
Text inversion embedings are usually faster but lower quality, but quite effective for style training (as opposed to subject training)

I'm a little out of date, so I could be wrong on some of those

[orchid dirge]

i want to generate background of this picture

but i dont know wherre to start

do i use painthua or get sd infinity/invoke ai>?

[runic osprey]

inpainting?

[knotty venture]

anyone knows how to make background cleaner? My generated images always has some random around the character. I've tried negative prompt: messy background, chaos

[wintry estuary]

Add noise, jpeg artifacts to your negatives and if you are using automatic1111 you can try adding a vae
https://huggingface.co/stabilityai/sd-vae-ft-mse-original/tree/main

[tired vigil]

/ok

[hardy trail]

anyone can help with

how do i prompt sd to give a very unsaturated picture

[silver valley]

hey someone knows how to get this style? or even just flat 2d citys backgrounds?

[blazing quartz]

synthwave?

[silver valley]

thx will try

[humble sedge]

So control net keeps spitting out nsfw results even when negative prompting. Never had issues with the models like deliberate before. Any ideas?

[silver valley]

describe the clothes more maybe

[humble sedge]

hmm might have been the model, stopped having issues when I switched to protogen

[still sigil]

Is there a way to make it so that the Automatic1111 gui will pick between some parts of a prompt for each generation? Sort of like what | does, but for entire images instead of each step?

this is for instances where I wanna just let it cook for a while, but I want a few different variations on a theme

[silver valley]

You can use x/y/z script with SR Mode as X

Select it and Hover over the name for more Information

[still sigil]

ok great thanks!

[slow tundra]

Hey folks, I'm trying to generate people faces with skin defects like acne, pimples, chickenpox. Have anyone tried to do so?

Chickenpox is a bit of exaggeration, but I want to have faces similar to what I can see in a subway, not in Hollywood under heavy makeup

I tried to mention keywords above, but all I'm getting is people with smooth skin with no flaws

I use ControlNet on sd1.5 right now

[silver valley]

Try intricate skin Detail, and maybe natural skin

[slow tundra]

Nope, doesn't work this way. At least for me

It feels like there is a strong bias towards photogenic people with flawless skin

[blazing quartz]

yes, since the data was scraped from the internet for image knowledge, and there are probably way fewer "pimpled face" people than models

[slow tundra]

Got GPU busy for a while, but it probably makes sense to try DB/LoRA with something like this

Gonna try it eventually, but yeah, considering that the look of skin defects depends on skin color also makes it non-trivial

[gloomy citrus]

Can someone tell me what I am doing wrong? I think I have everything set up right, but the images coming out, even with super simply prompts, are terrible.

[silver valley]

You need a resolution of 512x512 minimum because every 1.5 model is based on 512x512

[gloomy citrus]

that seems to improve things a bit, but still very poor results. I have an ok GPU, is that causing the issue?

I have a GTX 1660 TI on this machine.

[wintry estuary]

You have no negative prompt

Would recommend 512x512 euler a for sampler and go with 20-40 steps

[silver valley]

He also needs more Quality tags, also negative ones, also other models

[gloomy citrus]

can you recommend some good models? @silver valley ?

[silver valley]

Depends on your liking, Protogenv2.2 or Dreamshaper are good ones to start

[sage summit]

HELP, all the models i trained are looking great at 50-60% of generations and then starts to change to bad results, tried CFG an Steps but its the same, Does that have anything to do with over training??? i always like that range of my model but it starts looking old or changing to a different character.

[gloomy citrus]

ok. downloading both of them now... do I put those safetensors files in the same folder as the .ckpt files? or where do I put them to use them? @silver valley

Never mind, it is the same folder. This is much better. Thank you everyone!!

[cobalt sequoia]

is there a way to tell Stable Diffusion to only use certain colors? legitimately just giving it a list of hex/RGB values (for pixel art)

[tired vigil]

not currently, no. I tried to make it output any flat color and it was already hard to get the full picture in the wanted color with prompt

[proud gust]

which are the best samplers for photographic looking stuff? I mainly use EulerA and havent seen particularly better results from other samplers but not done exhaustive testing, what do ppl think?

[heavy holly]

i can use any image from here right? it's open source?

public domain i mean

[craggy steppe]

can some tell me what style is this? (img2img)

[pulsar coyote]

Idk if this is the right channel but I've remade the picture on the left, my version is on the right. Trying to get the same sharpness and detail and quality when upscaling and it's really close! But there is just some blurriness still in the image I've created (right). You can only really tell when it's zoomed in so it's cropped here but yeah. I guess I'm asking for the best settings for upscaling realistic anime images lol. On this I did combo of SwinIR_4x and 4x_AnimeSharp (0,5 visibility).

honestly it's probably like the details in the hair that I think are less pronounced in mine, like the lines in the braid...

[heavy holly]

hello i need help
I want to generate images that look similar to these above
what i want is that gradient lighting, and really smooth surface.

i have awful results so i need help

[lavish crypt]

Hello, I am trying to use inpant to change this figure to be explicitly hooded, but I think the "ukiyo-e" prompting is making it forcibly add a head to the image, could someone advise how i might go about encouraging the model to generate a hooded head rather than a visible one?

additionally, though not as important, I'd like to sharpen the image - either with the prompt or using an upscaler, if anyone had any advice on the topic I'd be grateful.

[pulsar coyote]

could try weighing the keyword like (hood:1.5) maybe, and maybe even do (ukiyo-e:0.8) if you think that's what's forcing the visible head to encourage it to make a hood!

[cobalt sequoia]

ah ok, ty

[autumn jewel]

@calm marsh this came out today thought it might help you with prompting a bit https://www.youtube.com/watch?v=HkLUmTJoyhw

[blazing quartz]

I personally use DPM++ 2M Karras. You may want to run the x/y/z plot and use sampler vs. steps to find something that fits your personal aesthetics.

[proud gust]

😀

[pulsar coyote]

are there are resources with ready to go presets(automatic1111 calls it styles) i would love to check out some good ones to learn prompt engineering a little bit better

[smoky totem]

yes i second the x/y/z plot with different samplers and steps. I find that depending on the model/ti/loras used, the results can vary. There is no absolute best. I use DPM++ 2M Karras for photoreal people often, but sometimes it is "too sharp" (hard to explain), maybe due to the lighting. I try that with environment/buildings too. D-2-K is sometime producing very unnaturally sharp edges, and worst, thickens details to achieve that sharpness, it is very easily noticeable when i just zoom in a bit to find that, dont even need to pixel peep. Whilst on Euler, things like fine mesh surfaces on buildings look "finer" with more detail, although it is a bit more "blurry", it looks more natural, and we perceive that surface to be finer with smaller holes on the mesh, vs D-2-K that made very sharp mesh, but with thicker mesh lines and much bigger holes on the mesh.

[atomic flume]

Anyway get better hands ?

(modelshoot style), (from_above:1.3), (hand_on_hip:1.2), (modern punk clothing:1.2), bulletproof vest, intricate design, 26 year old, (black woman:1.5), (vampire hunter:1.2), (she's holding a sword:1.2), (muscled body:1.3), (a character portrait:1.1), art by artgerm and greg rutkowski and magali villeneuve, analog style, (mdjrny-v4 style), (blood stains on clothing:1.4),

Negative prompt: lowres, (bad anatomy), (error body), error hair, ((error arm)), ((error hands)), ((bad hands)), error fingers, bad fingers, missing fingers, error legs, bad legs, multiple legs, ((missing legs)), error lighting, error shadow, error reflection, text, error, extra digit, fewer digits, cropped, worst quality, low quality, normal quality, jpeg artifacts, signature, watermark, username, blurry, ((error eyes)), ((bug eyes)), ((bad eyes)), bad mouth, error mouth, (error face), (((ugly))), (nsfw:1.2),

Steps: 50, Sampler: DPM++ 2M Karras, CFG scale: 7, Seed: 3286114148, Size: 512x800,
Model: protogenInfinity_protogenX86, Denoising strength: 0.7, Hires upscale: 2, Hires upscaler: SwinIR_4x

[blazing quartz]

hands are the bane of SD right now

could try inpainting, but not guaranteed

[tired vigil]

you can either inpaint again and again the hand, use a TI or LORA or other training specificaly made on hands to help the quality of those, or you could try controlnet and its openpose that has a hand model integrated it seemed

but yes, out of the box, SD hates hands for now

[blazing quartz]

I still get crappy hands with controlnet

I've almost given up and started just hiding the hands 😆

[tired vigil]

I only make turtles now, so I dodge the problem entirely

[tall wave]

trying to go for some abstract art that kind that replace a dancer's body with a pattern/ material(clockwork, clouds, grass etc), do you guys have any tips on models/ prompts that are a must have?

[atomic flume]

I wonder if future SD will fix the hands

[tired vigil]

double exposure maybe ?https://civitai.com/?query=double exposure

[tall wave]

let me give it a shot!

[winter ledge]

Hi everyone

How can i start creating? New to this totally lost

[blazing quartz]

If you have a decent Nvidia card, you can start with https://github.com/AUTOMATIC1111/stable-diffusion-webui

Or, alternatively, https://github.com/invoke-ai/InvokeAI

[terse python]

This one's amazing, really worth the try

[tired vigil]

real good job by @fresh nest 🙂 happy to share

[stone igloo]

@winter ledge

FAQ: I'm new here, how do I generate images ? Where is the bot ?

Welcome ! There is no bot currently to generate your images on discord. You may want to start by taking a look at the #1014939219904450590 channel. You can access Stable diffusion in different ways : 1️⃣ the official website, https://beta.dreamstudio.ai/. The easiest and fastest way to access Stable diffusion with 200 free credits. For any question on it, you can find help in the #1025467151206854736 channel. 2️⃣ Installing Stable diffusion on your computer. There are numerous projects that let you do that, and you will find help in the #🤝｜tech-support channel. 3️⃣ Running Stable diffusion in the cloud, through rented GPU services, using notebooks. You can find lots of them shared and discussed over in the #1011228442399883294 channel.

[tired vigil]

for a little more detailed answer, don't hesitate to ask

1/ is the easiest and fastest

2/ is the coolest and most fun to play with if you have the hardware for it

3/ is a cool fallback if you don't have the hardware for it

links given by happyfunball are quite good, the most popular tools for solution 2/

[candid dome]

Any tips on making a consistent character in SD? I thought making a Lora based on several similar results would work, but how do I GET several consistent design results?

[unique patio]

Not sure if you can do it with this tool but there are several options:

• Base the prompt on celebreties
• Train your model
• Use an image as input

.
.
.
Also got a question for help. I noticed some negtive prompts are using "::" can someone help me to find more info about this writting? I believe it is related to SD 1.5?

[silver valley]

Only ::? Or word:: ?

It shouldnt do anything

[wheat swift]

Yes training on generated images does work. it's called using a synthetic dataset

[idle tinsel]

This model is absolutely great when you just say a few descriptors and also I added a lot of comments on lighting etc (while it didn't turn that part out as I wanted exactly, still amazing. Like I gave a few ideas but this thing can think on it's own too tbf

Talking about models... This was supposed to be a model3 concept

[idle tinsel]

So I am having trouble reproducing effects/ambeince like the design behind (and maybe the placement of the car) in this picture. Scale is a bit off but it isnt a problem since the finishes are way better than any other image I have gotten it to produce. What should I have requested to get this again? Just in general, the lighting, colors. I mean I put this "... with surreal and majestic lighting, making it look heavenly. The ambiance looks dreamy..." but only this image came out with that effect

[tired vigil]

Hey! I've tried to make these characters look like me and my friends ffxiv characters. Did many iterations to get the cloting and body detailes kinda right. But after that many masking and inpainting the image is blurry and inconsistent. How can I clear the image without loosing detail/transforming the characters too much?

[tired vigil]

https://github.com/AUTOMATIC1111/stable-diffusion-webui/wiki/Features#prompt-editing

[to:when] - adds to to the prompt after a fixed number of steps (when)
[from::when] - removes from from the prompt after a fixed number of steps (when)

:: is a special operator that applies to the block it's put on, and puts a weight on the prompt presentation itself
In other words, the prompt :
a duck (on a plane::0.7)
will act like the prompt "a duck on a plane" for 70% of the image geneation steps, and then will act like "a duck". the rest of the prompt will not be presented to the model anymore during the last 30% of the generation process

this is a really nice feature

that I discovered yesterday, it would really deserve a prompt guide one day

[tired vigil]

I'm not 100% sure, but I would try to downscale this to 512x512 and then try different upscalers. some are very good with sharpness, some are good with color gradients, ... you may not find any that fits for everything, but you should find a better base.
Then you can use inpaint, and the "inpaint at full resolution" option, masking only small details and remaking only those. for example the hand on the far left, you would just mask that, inpaint it at full scale (it would do a full 512x512 hand and insert it in your picture), and keep on doing the same on the other details you need bettering

[tired vigil]

talking purely about prompt tricks, using keywords of context where you could see such pictures could help. like this could be a "car showcase" or even the name of a specific car event where good photos of car are usually taken. Those can bring good qualities to your rendering too
But mostly I would use some image2image or controlnet to force the composition, so I'd be sure to have the car in the center, the pillar, ... and then I could focus my prompt on the style and not that much on the composition since it would already been taken care of

[timid kindle]

Why does my nature landscape turn out so bright, contrasted and saturated even though I directly specified not to in the negative prompt and added some "high-quality" tags in the normal prompts (if you need more context to understand ask me)

[calm marsh]

Hi, the idea if the text2img result is close to what you like, you'd use that as an img2img to get more variation from it yes?
And if so, does the original prompt from txt2img has to be in img2img prompt? or you can start new?

[timid kindle]

It doesn't have to be there, no

You can use a separate prompt for img2img to more detailed description of the desired output

[calm marsh]

Thank you!

[silver valley]

Oh i knew the From when, but didnt know You can just use Word::0.8,
thx for the explanation

[river warren]

A trippy art piece with a celestial dreamscape portraying the solar system might feature a vast and expansive cosmic landscape. The stars and galaxies in the background would be painted with swirling, iridescent colors that create a dreamlike effect.

The planets in the foreground would be depicted as glowing orbs of light, surrounded by shimmering rings of gas and dust. Each planet would be uniquely stylized, with intricate patterns and designs that evoke a sense of mystery and wonder.

The overall effect of the piece would be ethereal and otherworldly, as if the viewer were floating through space in a dreamlike state. The use of vibrant colors and surreal imagery would create a sense of awe and inspiration, inviting the viewer to contemplate the mysteries of the universe.

[regal bobcat]

how to use image to image?

[tired vigil]

#1011634831467221033 could lend more responses but let me give a little sum up
img2img (image to image) is the art of modifying a picture using a prompt. It uses most of the same parameters as txt2img, but some new ones are important :

• the picture input. this will be the base noise of your new picture and, depending on the next parameter, will inspire a little or a lot your new picture
• denoising : this goes from 0 (the output would be exactly the picture you put in, and the prompt would be ignored) to 1 (the output wouldn't follow the input image at all, it would just use the prompt)

there are some other "modes" for img2img, in particular "inpainting". It's the same thing, but this time you let SD modify only part of the source image by drawing a mask on it

[tired vigil]

Hello folks,

Need help! Here is a great opportunity for a skilled Gen AI artist.

We are looking for a digital AI art creator with a keen artistic sensitivity to help us with prompt engineering. We are developing an app to help people get present to their emotions with the help of AI (https://FeelsArt.ai).

The artist will work with SD 1.5 and custom models to write many prompts with an aim to generate emotionally-meaningful beautiful art. Being knowledgable, attentive, accurate and with high emotional intelligence would be just terrific. When you  join our team you will be guided and well remunerated.

Do you know anyone who would benefit from this opportunity?
Cheers,
Nazar. 
[e-mail]

[tired vigil]

welcome around, nice presentation, and also a thanks for putting in the effort to explain it like that, clearly, and not spam every other day. I hope you got some contacts last time ?

[tired vigil]

cute cat too 🙂

[timid kindle]

Still relevant 👆

[steep jungle]

If I would like to get a littlebit "straighter", "cleaner" lines on my results, what are some prompts that would help with this, but not affect too much of the other stylistic stuff? (this is snipet example, 1 current 2 desired)

[tired vigil]

you may have tokens in the prompt that call for brightness indirectly ? like asking for a lion would force fur to generate too ? not sure, the full prompt could help (or could also stay a mistery, but we could try things)

[timid kindle]

I asked for gentle sunlight and visible sunrays, i don't think anything else was related. I deleted the prompt anyway now, I will try again later

[calm marsh]

Are there a good prompt for "axe" or "hatchet"?
I try to use inpaint to add an axe on the hand of the character. However I try the axe won't be there. I tried hatchet in prompt with no luck.
When I use prompt "axe" to generate new (txt2img) I only got something that looks like a short knife or sword.

[hazy yoke]

i second this problem~

[sick cipher]

Hello guys so i have a picture with a white blank background that i want to change do you have any tips of how to approach this ?

[vapid moat]

inpaint with a denoise scale of 1

[sick cipher]

What about the mask blur?

[vapid moat]

um probably can stay at default

[sick cipher]

Is this right ?

the inpaint options are a bit confusing for me

[vapid moat]

looks good to me

[sick cipher]

didn't work

could it be that the module isn't compatible?

I used AnythingV3

[unique patio]

@tired vigil thanks for the info, gonna try it out 😁, I saved the original text at home but I dont think it had any weights just the word and ::

[tired vigil]

I read it again later and I made an error in my example, please do use the doc I linked instead, I don't want to spread misinformation

happy to help :=)

[unique patio]

Jup gonna check the doc for sure.😁

[unique patio]

@tired vigil small follow up question (if you dont mind). What would you consider to be the best way of writing a negative prompt inside the actual prompt?

Should I just add a weight 0 or do you think about something else?

[tired vigil]

adding negative prompt inside the prompt is a bad idea.
weight 0 will make it ignored, it's also a bad idea since it will make no difference and still cost you tokens
some older implementations let you use negative weight, but I don't know how this behaves now. even back then, it was quite glitchy, outputing pure glitch picture more often than not.
Why not use negative though ?

I don't see solution directly to your question to be honest, maybe using tokens close to the "opposite" of what you want to exclude ? but using the token you don't want, in any way, inside the prompt itself, should only push it more inside than repulse it

[unique patio]

Yeah I am also trying to figuere it out. It is for a project where I only want to use pthe positive prompt

[smoky totem]

If it is only affecting 1 or very few areas, I would try in painting with a new prompt for the area, adding emphasis: straight lines. Sometimes it works. You may have to use a different sampling method, from my past experience. However, I would also like to know if there is a better way, more consistent way to avoid these wobbly lines in Stable Diffusion. I've used positive prompts like straight lines, sharp lines, fine lines, negative prompts like wobble/wobbly lines, curve lines/shapes, badly drawn, scribbles, etc. There is never consistent results, and like you say, sometimes they affect areas where I want curves, so it is not a good solution. Other times, SD just straight up ignore my calls for straight lines, drawing wobbly lines all over the image, despite everything is based on rectangles, and end up drawing a whole page of soap like bars.

In those cases, in-painting wont work since I need SD to completely redraw it. I've tried image2image to try to save the work since I like the color, lighting, composition, but that does not work 99% of the time. I will draw something new to get straighter lines out, and if I control it so it says true to original image, the soap bars reappear without being straightened out.

This is one of the most frustrating problems in SD. switching models, aspect ratio, sampling method, prompting, nothing fixes it. Either get a completely new picture for straighter lines, or live with the soapy looking images.

[smoky totem]

case in point, i cant in-paint my way out of this mess...i like everything about it, except the wobbly lines

again, verticals are straight, but the horizontal grid/mesh, wobbly

[bleak fractal]

Is it just me or is stable diffusion not good with generating a lot of basic items such as sword, staff, pickaxe, pen, hat, arrow, bow, crossbow, etc. I’m guessing it hasn’t been trained on a lot of things? Everything like this that I try generating gives me disfigured objects or very zoomed in close ups that don’t resemble these items at all.. is there a way I can get these to generate correctly or somewhere I can report this bug?

[smoky totem]

SD does not like violence. we talked about guns a few days ago

you may want to train your own loras and name the bows axes etc into other names

[bleak fractal]

I could imagine guns but a lot of this stuff might be for recreations of the past like medieval times

[steep jungle]

@smoky totem Thank you for comprehensive response, appreciate it. I will try to remember to let you know If I find something that helps with it 🙂

[tired vigil]

Need somebody to help me w/ making an image to a prompt (kinda confused how to replicate the art style, dm me (NSFW))

[idle tinsel]

Thanks! Can you make me a simple example of any car you want its all ok

[idle tinsel]

For NSFW what I do know is u need special models or something idrk

[tired vigil]

sure, let's try those tips, but I'm not into cars that much personally, I'll be worse on prompt itself.
1/ using CLIP on your pictures, I got some tokens and added mine to get the prompt

a silver car on display, car showcase, square picture, Dahlov Ipcar, ue 5, a digital rendering, panfuturism
first picture as result. problem is, it's not consistent. I get around 6/8 cars in frame still, quite OK to me
there is no real lighting, and the "car showcase" token put it inside if I don't add anything for describing the outside
But also, using the first pic I sent in img2img with 0.75 denoising and some new tokens like "surreal lights, lens flare" and adding "outdoors" to the display at the start of the prompt, I get the grid I sent. Lots of consistency, cars stay in frame, lights get better.
2/ let's start again with the black car this time, using controlnet. I used interrogate to have a new base prompt and added kind of the same modifiers :
a black rolls royce parked in front of a building, Andrew Law, luxury, photoshoot, car showcase outdoors, surreal lights, lens flare
this time, we get the grid I sent. Like said, there is no more question about composition : the composition stays exactly the same as the original photo, but you can play with all other parameters now and get it just as you'd like.
This was done using the "canny" model of ControlNet

all on 1.5 model

[idle tinsel]

I don't get the majority of it, my appologies lol. Im basically only skilled enough for the basic playgrund on the online versin (also hardware constraints) so basically do you have any tips for me on just using that?

[tired vigil]

mostly the 1/, I'll rephrase
useful tokens in my tests : outdoor car showcase, square picture, digital rendering, lens flare, surreal lighting
You can also use the "Image" field in the bottom right, and put an example photo with the composition you want : a car in the middle.
When using that option, you have a slider named "image strengh" that appears at the top. lower it a little, like 35%. This is how strong your input image will inspire the output. If it's too high, it will have a hard time changing too much things, but you can play around with it.
Using the prompt :

a black rolls royce parked in front of a building, Andrew Law, luxury, photoshoot, outdoor car showcase, square picture, digital rendering, lens flare, surreal lighting
I got this :

[pulsar coyote]

from automatic1111 github: "Adding a BREAK keyword (must be uppercase) fills the current chunks with padding characters. Adding more text after BREAK text will start a new chunk." what does this mean? whats the use case?

I suppose it might be to do with doing prompts like :girl riding a bike BREAK sunshine, street, cars. So seperating concepts and foreground and background . This is what I want it to be though I don't actually understand the explanation lol.

[tired vigil]

check the thing just before it in the doc, Infinite prompt length. it's made so that you can manage "batches" of token in their weird way of making infinite prompt possible

Typing past standard 75 tokens that Stable Diffusion usually accepts increases prompt size limit from 75 to 150. Typing past that increases prompt size further. This is done by breaking the prompt into chunks of 75 tokens, processing each independently using CLIP's Transformers neural network, and then concatenating the result before feeding into the next component of stable diffusion, the Unet.

For example, a prompt with 120 tokens would be separated into two chunks: first with 75 tokens, second with 45. Both would be padded to 75 tokens and extended with start/end tokens to 77. After passing those two chunks though CLIP, we'll have two tensors with shape of (1, 77, 768). Concatenating those results in (1, 154, 768) tensor that is then passed to Unet without issue.

[pulsar coyote]

oh okay, so it needs some help breaking it up. Is token= single word or is token what's inside two commas? ,girl riding a bike, = 1 token, or 4?

[tired vigil]

it's a little more complex than that... a token is around 3/4th of a word usually, and is broken apart automaticaly by the interpreter there. You see the number of tokens grow as you type your prompt in the UI too

I can't check right now, I'm running a training, but your example could be around 4 to 6 tokens maybe

this also can depend on some other things, like if you had a TI loaded that would weight 8 token and would be linked to the word "bike"

[pulsar coyote]

"it's a little more complex than that" it always is haha! Easy enough to keep track of it though in the UI. Ty!

What does a colon and underscore do to a prompt? Ive looked through the automatic1111 doc but didn't see it mentioned. I've heard some explanations for it like ":" can link two concepts like so : cat:girl. But I see no mention of that. I've also seen underscore described as a way to limit a keyword to only affect another keyword like so: white_dress. And the explanation was that it is done in order to prevent the keyword white from "spilling" all over the image.

[unique patio]

I got a special model, if you have a specific request I could try to run it once, but cant promise anything. Just for comparision.

[queen kiln]

Question how to prompt color or not just red,green etc simple colors is it possible like hex codes?

[tired vigil]

you can try to trick the AI into it but there is no real command for it, you'll ask for a prompt that would describe such things. Using terms like the HEX code, the pantone name, .... copying how such picture could be titled if it was found online, can help, and did for me to some degree, but there was no consistency, I could get a picture of a pantone book in the middle of flat colors... this isn't the best tool for that.

[tired vigil]

I realise that

[pulsar coyote]

testing diffrent samplers in img2img and the image looks normal while rendering but then at the last second turns to this mess on the right

whyyyy

[tired vigil]

Dm me

[tall wave]

afternoon! been trying to recreate this style in SD img2img but i cant seem to figure it out
tried a mixture of prompts and models but no where close to what i want. anyone got some insight on how to achieve this?

[runic osprey]

i cant even describe what im looking at so idk how an ai would

[tall wave]

this is it

https://www.instagram.com/p/CeoXYqFKbVS/

miscsoul

Modern times flow.

When dance meets AI.

Performer:
AI artist:
Music:

Likes

81224

Instagram

[unique patio]

how about: a black and white photo of a person made out of metal parts, cyberpunk art by Steven James Petruccio, cgsociety, afrofuturism, dystopian art, steampunk, future tech

Credits to image2prompt AI model

[timid kindle]

This happens to me a lot. The image looks great during the sampling but the last sample turns out too sharp and saturated, just like your image.

[indigo trail]

Is there some kind of tutorial on the basics of creating solid prompts that produce okay or decent results that you can use to refine going forward?

[indigo trail]

how do you even get to this level of quality in the first place?

[weak kite]

usually, a really good model

maybe this is prompting help, so i'll ask here

i struggle pretty hard to get ControlNet to work
it functions, it copies the pose of a source image, but the generations that come as a result of it are invariably awful

[pulsar coyote]

The model is neverending dream made by Lykon and yeah model+prompts lol https://civitai.com/gallery/114071?modelId=10028&modelVersionId=11925&infinite=false&returnUrl=%2Fmodels%2F10028%2Fneverending-dream-ned check the comments - the creator kindly explained his workflow to me.

But one trick for detail is in this video https://www.youtube.com/watch?v=4u-Ytioi3DM&t=1s. Result is like this. Obviously the left face has TONS more detail.

[indigo trail]

it's a safetensors file, does this work with SD?

[weak kite]

yeah, safetensors work

the pruned versions of ControlNet models are safetensors iirc

[pulsar coyote]

yep with automatic1111 it works

[indigo trail]

so I just save it into the stable-diffusion folder like all other models?

[pulsar coyote]

exactly the same

[indigo trail]

I'm excited to see how this model differs from the default

[weak kite]

not sure if this workflow is applicable for people using computers without a ton of ram

i'm using a 3070 which is 8GB, i cant render above 1024 in any dimension

[smoky totem]

I've experienced this, what I tried to fix this is lower/upper the CFG little by little. Usually within like 2 CFG steps the good image shows up. I am not too knowledgeable on this but the image on the right that is screwed up is either over/under "baked" (edit, sorry not CFG but steps, I usually start with 20, but most models can produce "acceptable" images from 15 steps, then you start to see pose changes every 3-5 steps. Sometimes I watch the scenes change from 20-50. If the eventual image looks bad, I step back a couple of steps or forward a couple of steps, usually I'd find the image that I want.)

I would capture the prompt, seeds, settings, and redo that image with small decreases/increments

[pulsar coyote]

I'm on a gtx 1080 and I'm ok starting generating at 512x768 then doubling it and then inpainting into that.

[weak kite]

weird, mine always hits me with a CUDA error
you running optimization code or anything?

[pulsar coyote]

nope just standard everything. But yeah that happens sometimesbut as long as i stick to that res I'm almost always ok. If it absolutely refuses a reboot does the trick but it's rare that I need to do that.

[weak kite]

lemme give you a life example for ControlNet just sorta shitting itself for me

[pulsar coyote]

But slightly higher res and it's CUDA errors all day

[weak kite]

i want to use this D&D char ref as a pose

[pulsar coyote]

easy enough, you'd think!

[timid kindle]

How did you even get this clothing so similar? I understand how one can take certain key features of an image and sort of replicate it, but I could never get a specific type of clothing like that

[weak kite]

i dont exactly want anything funky so the prompt is basically just "greg rutkowski, elf"
cause i just want to see a re-rendition of the image in his style

instead it gives me this

[timid kindle]

So do you want to only save the pose?

[weak kite]

my intent is to have the same image but to have basically changed the art style, i suppose?

but what often happens is i just have the "outline" of the pose remain, while everything else just becomes random colors

[timid kindle]

Interesting

[pulsar coyote]

I just copied the generation data. I had to generate some images to find that pose though. But since that image used a Lora of the character it pulled from a kinda limited dataset. BUT what I gather is that same generation settiongs = same image IF on the same machine. Like when you press generate and then generate again - same seed, same image.

[weak kite]

even when i lower the CFG and even the weight of ControlNet, it still gets obscenely funky

[timid kindle]

Ah, alright. I can't imagine making up a prompt to replicate the original image so closely without yoinking data from somewhere.

[pulsar coyote]

Ok so the colors will not transfer by controlnet. What contorlnet model are you using for this? If you give me the prompts i can try it out.

[weak kite]

dreamlikeart, greg rutkowski, elf, beautiful, handsome, male, masculine
Negative prompt: feminine, woman, female
Steps: 20, Sampler: Euler a, CFG scale: 3, Seed: 552418222, Size: 512x768, Model hash: 0aecbcfa2c, ControlNet Enabled: True, ControlNet Module: canny, ControlNet Model: controlnetPreTrained_cannyDifferenceV10 [ea6e3b9c], ControlNet Weight: 0.4, ControlNet Guidance Strength: 1

the model is dreamlike-diffusion

[timid kindle]

Now I feel pressured to look at different models, because I use automatic111

[pulsar coyote]

i do not have that model but ill try with another one see what happens

[timid kindle]

Wait sorry I'm interrupting, but for example in the automatic111 webui, are the LMS, Euler A, etc different models? What are they?

[weak kite]

sampling methods

[pulsar coyote]

ok so I would go up on the weight in contorlnet. And it needs more prompts I think, the AI doesn't 'get' what those clothes should look like. But your image look almost like a VAE issue - try a diffrent one or none at all maybe?

[weak kite]

weirdly enough i got better results by dropping the weight to about 0.4

not great

but workable with some inpainting

[pulsar coyote]

i forced him to wear clothes but it took some additional prompting

[timid kindle]

The left hand looks interesting

Huh, actually both hands

Even ai has tough time drawing hands

[pulsar coyote]

interesting hands - the stable diffusion classic

THIS one got BEAUTIFUL hands on accident

[timid kindle]

Wait this is pretty fire

The fingers look as if they are connected but it's almost perfect

[pulsar coyote]

They are close enough that upscaling would fix it I think, it's still a pretty small image

shockingly good for low res like this

[timid kindle]

It seems it's much easier to generate landscapes and nature than humans and objects. I wonder why

[shadow spruce]

how do i manage to get SD to give me more detailed texturing? surfaces like skin and clothing tend to come out flatter for me..

[short moon]

I am really having a lot of trouble with controlnet: it just mangles faces when the character is not immediately in focus and it's driving me crazy. Is anyone else having a similar problem?

[smoky totem]

inpainting to the face afterwards

https://www.youtube.com/watch?v=MDHC7E6G1RA

5minutes in

[short moon]

thank you so much. I realize how little I use any tab except txt2img

[smoky totem]

you are welcome, for more advanced control on how to fix faces using inpainting, ask someone else here. I rarely make portraits of characters

[hallow elk]

Hi guys, how can we get the past prompting that we've made?

[tall wave]

does stable diffusion pref working with models agains a white or black bg?

finding that im getting weird artifacts when isolation a subject on a black background

[lone badge]

you can drop the image made inside the "PNG Info" tab in automatic, or look at the image metadata. If you didn't change the options, the prompt is saved there with the other parameters, like the seed

[hallow elk]

Thank you! Will try for that!

Perfect! I got it!

Can we use parameters in stable difussion? Like in midjourney, exampel ::1 ?

[lone badge]

there are lots of parameters and options yes

you use Automatic from what I got

check that wiki, it"s really well done page and shows most of the features. https://github.com/AUTOMATIC1111/stable-diffusion-webui/wiki/Features
In particular, some of the features pertaining to prompt edition

• https://github.com/AUTOMATIC1111/stable-diffusion-webui/wiki/Features#attentionemphasis
• https://github.com/AUTOMATIC1111/stable-diffusion-webui/wiki/Features#prompt-editing

[stone igloo]

FAQ: What is Stability AI?

Our vibrant communities consist of experts, leaders and partners across the globe. They are developing cutting-edge open AI models for Image, Language, Audio, Video, 3D and Biology.. AI by the people, for the people. Learn more here

[hallow elk]

Thank you so much!

[lone badge]

you're very welcome 🙂

[topaz wharf]

I had a lot of help from the PromptHero website in the beginning understanding the Parameters and Options by checking out the stuff there on the website and what prompts were used to achieve those. Check it out maybe it'll be of help.

[kindred garnet]

i always get bad eyes in my disney artworks -> they look more like human eyes

any idea how i can fix this?

[barren meadow]

have you tried prompts for simple/cartoon/disney eyes and/or negative prompts for human/detailed eyes?

[kindred garnet]

i might try this again, in general my eyes look always similar, also in different anime styles, it always feels like i have a specific style with a humanized face

-> in other models too

[barren meadow]

I have a question on attention/emphasis I don't see in the FAQ: can you nest groups? For example, (a (tall man:1.5) with an (umbrella:1.5):1.5)?

[full timber]

Any similar webs like prompthero?

[errant perch]

is there a way to tell if the model you're using is trained with a given tag?

[vapid bloom]

I'm having some trouble building a prompt for a description I've written. I can get close, but nothing quite matches the "image" I'm envisioning in my head. can anyone help me out or give me suggestions on how to convert the following description into a coherent prompt?

The city is a labyrinth of towering, concrete buildings, their facades marred by the scorch marks left by years of industrial pollution. The streets are lined with rows of identical gray buildings. The sky above is a sickly shade of yellow, and the air is a thick smog with the stench of chemicals and pollutants. One building stands out: Nutri-Synth's headquarters, a towering structure that looms over the city like a monument to greed and power. The streets are filled with robotic drones, carrying packages and delivering synthetic food to the masses.

[south solar]

How would you describe a dryad as having bark for skin? I have been trying "bark on skin" "wooden skin" and such, but not much luck. Or is that something that would probably require an embedding?

[wheat swift]

skin off bark and moss worked when I did it

[south solar]

Thanks. I will try that.

[south solar]

Took some work and a bit of img2img, but I now have a portrait for one of my rimworld pawns.

[torn hawk]

Hello~ I have a question about the resolution. I am using other people prompts to create an image, it's ok with low resolution such as 768 with 1024,etc. But when I am using around 1024 with 1526,etc or higher, there'll be extra legs occur with image. Do I need to install a plugin something to solve this? Because it's actually others prompt with no problems. Thanks if you give me a helping hand

[smoky totem]

upscale after you generate the initial image if you want a higher res image, use outpainting if you want more background to show up beside your portrait subject, do not just up the resolution from initial generation. all kinds of floating limbs and weird stuff will show up if you go too far above a model's initial training resolutions, most are 512 for sd1.5 or 768 for sd2.1 (edit spelling)

[hallow elk]

Will do! Thank you so much! ⚡️⚡️⚡️

[white lily]

hello,I want to use the original design image of my product with a white background, without any changes to the product's design (including style, color, details, etc.), and have an AI generate different background images of the product in various usage scenarios or angles based on my textual prompts. It is important that the product in the original image remains unchanged. Please contact me if you can fulfill this request.

[karmic bloom]

Is there some documentation or maybe a youtube video somewhere that gives an overview of the ins and outs of prompt generation? I see prompts posted with things like (monochrome:1.3) [out of frame] (((extra fingers))) and I'm curious to know more about them.

[sacred swan]

how can i prompt on google collab notebook v.05. I dont see changes in animation , and do I need to rpompt everyframe

[regal bobcat]

thank u @tropic topaz

[lone badge]

you are welcome 🙂 (Aaron is my currently unavailable account, sorry, it was named Guizmus)

[glossy basin]

Random Seed a Batch of Identical Images
I'm using Automatic1111 stable diffusion webUI. My goal is to take a still image and create a sequence where it cycles through different AI treatments, basically a different seed per frame. I was hoping to avoid the manual task of generating all the images separately and sequencing them in video editing software.

I was hoping I could do this by making an image sequence that consists of the same image each frame and then using Batch img2img. However, I'm finding that there is very little (if any) variation between frames, even though within the Extra seed settings with a Variation seed of -1 it made me think there would be a different seed each frame.

Why is it the case that a batch of 4 images in img2img can turn out so differently, but my frames in Batch img2img are coming out so similar? I'm obviously missing something. Please could someone point me in the right direction? As you can probably tell, I'm extremely new to all of this but keen to learn!

[kindred garnet]

anyone has an idea to get better eyes? always when txt2img, my eyes look not really great. messed up. even when i rework them. Im using classic anim atm and trying to create classic disney stuff.

as an example

even worse

[silver valley]

you dont use any negative tags, thats your problem here, try add stuff like:
blurry, low resolution, deformed, bad anatomy, etc to the negatives

this is with my negative tags, same model, tags and steps,

[mental shadow]

I'm struggling a bit with inpainting... I'm trying to patch up this little spot of exposed skin which should be the black dress.
I'm adding "black dress" to the front of my prompt, and "skin" to negative. Masked content: Original. Full Picture. I've tried various combinations of denoising strength + CGF scale
I must be doing something blatantly wrong, just not sure what that is yet 🙂

[kindred garnet]

oh ok , thank u 🙂 -> is it possible to read any info from my pics? 😄 haha didnt know that. still looks a little too humanlike, but the bird looks great in general, way better than mine

[silver valley]

yea generated images contain exif meta data with all infos you used to generate 😄

[lone badge]

ok so the problem here is that img2img morphs pixels, and it can be hard to go from full flesh to dress like that.
Open your source image in paint, and add that black spot to it too. then go back into inpaint, put a mask over the bad black you added, and you'll make it a lot easier to the AI to modify it

[silver valley]

you can load your image in PNG Info tab in the webui to see all settings

[kindred garnet]

thank u so much for the advice. im really completely new to everything. in general when a model is based on something is it literally like SD 1.5 with an addon? or how does it work?

im wondering i can generate similar images with models that are like 1 GB and specific for something and then i have those big 10 GB ones. Are they just trained with more stuff?

[silver valley]

there is SD 1.5 and 2.1 as base official models, now the community can train on top of this base with their own images (fine tuning) to get better results. So it has the core model + extra trained stuff in it

[kindred garnet]

nice, i see 🙂 thank u

[silver valley]

All models are over 2gb, if its smaller then its mostly a lora or embedding, these are like (additions to a model that go on top)

[mental shadow]

Aha - thank you very much for the tip. I'm on track now, I think I just need to fiddle around with the sliders a bit to get it perfect now

[lone badge]

or hypernetwork too, I know they aren't the most popular right now but I still love them

well done !

(just had the time to see)

[mental shadow]

This was the end result 🙂

yep just painting a black blob there before inpainting did the trick

[acoustic ivy]

hi, I am using automatic1111 webui. What does < and > tags do in the prompt? like <cinematic light> thanks

[atomic flume]

does word case matter to SD ?

[silver valley]

they do nothing, here are more information how to use () or []:
https://github.com/AUTOMATIC1111/stable-diffusion-webui/wiki/Features#attentionemphasis

[silver valley]

what do you mean?

[atomic flume]

Upper or lower case

for words

[silver valley]

it wont matter, it just reads them, but it could be that it gives you different outputs for Tree or tree

[atomic flume]

also should I use | rather than commas ?

[silver valley]

no you need , for seperating. | is used for other stuff (Alterning Words)

[atomic flume]

Ah thank you

[steep aurora]

Trying to start using Loras more. I frequently get this 'overbaked' look that occurs, and it seems like it's there even if I turn the strength down on the loras, Wondering if the total lora strength needs to add up to 1, or if I can have 2-3 loras turned on at high strength levels like you can for embeds?

[karmic bloom]

any tips for improving blending of seams between the original image and inpaint-sketched generated content? I'm using auto1111 for reference.

[steep aurora]

I've had some success with increasing the 'blur' up to 20 or so. Also decreasing the denoising works pretty well. Usually 0.4-0.6 is what I use for getting rid of seams. Unsure exactly what blur does, but it seems to work in blurring seams for me.

[dark current]

Nbbnhj

[kindred garnet]

Can someone explain me the difference between lora and embeddings?

[narrow needle]

2 vagina

[trim ether]

any tips to make 1male and 1female with specific characteristics each?

[quiet mesa]

use inpaint

[gritty summit]

Does anyone remember a MJ prompt guide that had a ton of resources sorted by style, lighting type, pose setting, etc that had a bunch of examples in each categories drop down? I SWEAR I had it bookmarked but I can't find it anywhere 😦 Open to other suggestions if you have one

I was using it on SD and it was super helpful for prompt words on the scenic set up of the image

[smoky totem]

your aspect ratio is way off for this figure. shorten the height by a lot to fix this. otherwise your fig will either stretch or 2 bodies show up (even if you neg prompt to control those). If you want higher res, upscale afterwards or use highresfix during generation. if you want more background, outpainting afterwards, do not just raise the resolution

[noble sigil]

do you want to have entire legs visible? are you using prompt for full body?

[balmy glade]

I'm using the stabiliy API img2img. I would like to set one of my pictures as input (init_image) and edit that picture (prompt) so that the output will be that picture of me wearing a golden armor or a suit, or with another picture style. But the face of the person in the output image is totally different of mine.

I don't know what I'm doing wrong !

What is a better way to keep the face of the person in the init image. I would like to create avatars using the API.

[lone badge]

[fair girder]

Hi👋🏼
Any idea who to get really empty backgrounds with nothing but white? Tried prompting it but it only works about 40%.
Impaint does work, sure but that's just for one picture, i need it to work in generation

[lone badge]

you could use img2img with a very high denoising, like 0.8 or 0.9, on an image with just a white background and a basic stickman, or sketch of a person, maybe colored in a color you like.
since pixels get morphed during that denoising process, they start from that while background, and won't move far from it since the prompt doesn't allow for it. As for the pixels of the stickman, they'll change quite a lot. The base picture will just be a "composition guideline", pushing in the good color patern as input noise (input image is input noise) and the good picture composition through that stickman.
If you want to control more the composition, like the pose or even some details of the character, you could also use controlnet, but your question seems to indicate you want variety, so I would go with the method I described
ask if I wasn't clear, there is a lot of terminology in there

[fair girder]

So basically a prompt saying "just use 500500 of the 768768 isn't possible? Will try your approach, thanks

[lone badge]

yep, you cannot just control the pixels color through the prompt. it's really not thinking like that.
It was trained using how pictures were described/named on the internet in part. so try more to think of a prompt that could describe the image you want if you found it on a forum for example

like, character on white background, I'm thinking "character sheet" would be powerful

you need to find examples in your mind of how you would find your image "in the wild", and describe it how it would be described

like, when doing realistic photo, you can add "taken on iphone 6" or "70mm" and it will work quite a lot, pushing SD to make only realistic looking photos

same goes here

[fair girder]

Hm, will test around more later. Anything background helps but not good enough. Not touching border did nothing at all. Clip art just a tiny bit

But won't i lose the randomness wirh img2img?

[lone badge]

with very high denoising, no

if you had 1.0 in denoising (don't do it it crashes) it would be the same as if you didn't use any input image

so there is a right spot, usually around 0.85 for me, where the input image is no more than a suggestion for the composition

you still may want to check the example uses of controlnet if you didn't yet, one of those could be of help, like Segmentation or Open pose

[fair girder]

Sounda promising, thanks. 3 more hours😅

[lone badge]

this is txt2img, the image on the left gets preprocessed to only use the composition, not the colors or details, depending on the mode you use

https://github.com/lllyasviel/ControlNet

[fair girder]

I got CN. Didn't do as i wanted for this issue. At least mit yet

[lone badge]

i'm still learning how to use it correctly, so I stay a little shy giving tips on it tbh

hard thing, lots of possibilities

[noble sigil]

Don't paint over your face in the img2img

[fair girder]

We are all new basically

Anyway, will test later, thank you

[slim orbit]

Hey

Guys

Need help with prompt

[balmy glade]

How can I do to avoid painting over my face ?

[noble sigil]

When using mask just don't use it on face.

[balmy glade]

It means a mask for each image.
There is not a way to automate that ?

[noble sigil]

I suck at explaining, watch this https://www.youtube.com/watch?v=3QPt_dPnmGk&ab_channel=Colinfreeman

[astral fern]

hi, i connected the diffusion.gg bot to my discord serwer, but when I try to draw it keeps saying that "the app doesn't react"

[severe phoenix]

Hi guys, how can i get rid of texts as much as possible when generating? Trying with negatives (watermark:1.2), (logo:1.2), (barcode:1.2), (UI:1.2), (signature:1.2), (text:1.2), (label:1.5), (error:1.2), (title:1.2) but they still generate

[formal ember]

im still looking for guides for pose prompts, stuff like angles etc, are there any public guides for it out in the internet ?

[noble sigil]

https://www.youtube.com/watch?v=ZCJX5ZAk9SA&list=WL&index=64&ab_channel=OlivioSarikas I haven't tried it myself

[noble sigil]

I only use (words) (text) (copyright)

[tired vigil]

What command create pics?

[noble sigil]

Any, you can even get a pic with empty prompt

[tired vigil]

Why it send me to this channel instad of prompts

[silver valley]

@formal ember here some views:
https://www.reddit.com/r/StableDiffusion/comments/y46k9t/comment/isd3xqd/

[native idol]

Hello everyone, I'm Vega, I'm an animator that's interested in learning Stable Diffusion and all of it's features. Are there any guides anyone would recommend? My goal is to use my own art to feed into a style on top of an animation through image sequences. But for now I want to learn how to use the features and the basics.

[lone badge]

hey and welcome around Vega 🙂

I don't have a good guide, but there are multiple resources and paths I can point to you, to discover SD. Each path can take a long time and have lots of other subpaths, it's like an hydra, it feels like you can't really learn everything, but keep cutting heads, one after the other, and you'll progress for sure.
1/ there are lots of tools to use SD. Some can be installed localy and give you more freedom to explore it without limits that you'll find online. They can still be used online if you can't run them on your computer. The main two people talk about would be
1.1/ Automatic1111 (https://github.com/AUTOMATIC1111/stable-diffusion-webui/)
Lots of features, lots of sliders, lots to learn. There is a quite good wiki that can also be a good thing to check, all features (almost) are showed and give a good idea of what you can do in SD. (https://github.com/AUTOMATIC1111/stable-diffusion-webui/wiki/Features)
1.2/ InvokeAI (https://invoke-ai.github.io/InvokeAI/)
Less feature but very high quality UI with a focus on inpainting/large canvas. here is a good example use (https://www.youtube.com/watch?v=IuJv4EMFq1s)

This first steps should give you a good idea of the possibilities. Then choose one, install and try.

2/ prompt making. It's a real thing to learn, it takes time and experiment, and asking around here, or using some prompts you find online on sites like https://lexica.art/ to learn from. Also, chatGPT is a great friend when it comes to making good prompts now, I put an example on how here: https://discordapp.com/channels/1002292111942635562/1002292112739549196/1079658399366643712
another guide was published by one of our mods, and is pinned in this channel, very nice explanation of one way to make prompts
https://discordapp.com/channels/1002292111942635562/1011743094309396631/1030121511845101638

3/ training. You can train lots of new things into your IA model, teaching how to make things it didn't know, or refining concept it did know of. This can be very efficient in animation, to help keep stability from frame to frame. Here is a guide on style training. subject training is a little different but very close https://github.com/nitrosocke/dreambooth-training-guide

There are lots of other things, especially pertaining to animation since it's what you are interested about. In particular, the tool Deforum is specialised in animations though SD : https://deforum.github.io/

Lastly, the big game changer that came around recently is a complement to SD, conditionning the output to help you control them, to help "tame the beast" that is the model and make it give you the intended results. It's name is "ControlNet" and it's very powerful for animations too. https://www.reddit.com/r/StableDiffusion/comments/119o71b/a1111_controlnet_extension_explained_like_youre_5/

WALL OF TEXT POOOOOOWAAAAAAA

[native idol]

wow

this is so helpful

thank you so much

[lone badge]

I reformated a little

and added more guide links

no problem, I can link other people to it in the future, it's a common question and there is a lack of detail on it around, easy to access

[native idol]

ControlNet sounds really cool

[lone badge]

it really is, I'm having so much fun with it

like... forever alone guy

[native idol]

ahahah

wow

memes are going to be the next level

[lone badge]

https://new.reddit.com/r/StableDiffusion/comments/11e4dfw/ragecomics_and_memes_as_real_as_i_could_get_them/

my "work" of today lol

[native idol]

so could I use my art as a means to using a style?

like my idea is to animate in 3d, use my paintings as a style

[lone badge]

you could yes.
Either you could use you art as base, and have SD change the style on it
or you could teach you artstyle to sd, making it understand "painted by Vega" in the prompts you would then be able to make. And it would mimic your art, if you managed to train it well.

[native idol]

how many images do you think I would need? 100s?

[lone badge]

I trained some styles, on specific animes, videogames, ...
I would usualy say a style needs at least 30 to 50 pictures, but will benefit from more.
The important thing is to keep it diverse. If you repeat the same art or composition too much, it may pick on it too much too, and put it in all the outputs.
Last style I trained https://civitai.com/models/1158/mosaic-art
I used 46 pictures. I don't have examples of lots of types of subjects, like for example I have very few "landscapes" or "sky shot" in the dataset, and because of that, the model is not very good at doing those.
So the more diverse subjects you have, the more your style will adapt to any prompt you give it
But if you have a very narrow style, you still can train. it won't need as many pictures, and it will just be able to output in your narrow style

it all depends on what you want to target

training is really taking the Pot in witch SD already is boiling, putting fire on, adding small pieces slowly and baking a new model

you need to know though that the model doesn't "grow", it will forget things, become worse at things you are not training it on

it's "specializing" the model, so it's important to keep the limits in mind

(that being said, there are numerous training methods, and some can purely add new data in some kind of post processing too)

[native idol]

https://twitter.com/OmniVeiga/status/1630243261670658048

Vega (@OmniVeiga)

Paintings!! Just for some who don't know I also use to paint a lot back in the DA days. I'm hoping to finish something soon while working on animation projects.

Twitter · Feb 27, 2023, 4:27 PM

here's a small sample

so would you say as diverse like this?

[lone badge]

something in this style yes.

I would be very careful in how I train on that though

the pokemon ones in particular

[native idol]

more humans then

[lone badge]

pokemon token is quite powerful in the model, it can mess things up

no, you can go monsters too

but taking something like yoda or pikachu can pull on weights in the model that are quite strong, and that can mess up the training

[native idol]

oh since it's such a huge promt

ok

[lone badge]

basically, your style would get trained faster on a random thing than on a very strong thing

[native idol]

so avoid fan art?

[lone badge]

or go full into it in one category, and you'll train enough to "cancel" what was in the pokemon token

but don't go half Strong token + half weak tokens

I did that error with the manga death note

Riuk is very well trained already, some others aren't

I couldn't get every character trained correctly at the same time in the end :p

I should try again

but, like, with just a few pics, it's hard to say, but even if I had your full dataset it would be hard.
I have made around 20 quality models now, and I can't know for sure before I tried, before I checked the training curve, before I tested the first trained model to see what went wrong. And then I would come back to my dataset and change pictures

and do it again until I get something of quality, that isn't all pokemon because of 3 pics, or that isn't all blurry because I made one pic to blurry, or hasn't some text (try to NOT have watermark in your dataset), ...

lots of biases can show up, and are hard to anticipate

it comes with experience

(I do have 2 friends that have made that their real job, this is a real complicated subject)

[native idol]

I can't wait to start playing with it tonight

[kindred garnet]

someone have an idea why my faces always look more 2D especially in more 3D anime mixed stuff.

i always have a body with more depth and it looks like a 2d face is set in

[formal ember]

thank you, i will look into these links

[fair girder]

hey, back again.
That did kind of the trick.
But i'm using https://github.com/KutsuyaYuki/ABG_extension.git now as well. works great in 90% of the time.
Thanks for your feedback, really did help me

[rain jay]

what should be the prompt to make her face this way?

[mossy cloak]

Hey folks, I'm a software engineer working on a app for professional prompt engineers. I'm looking for a few people willing to try out some early versions and provide some feedback. I have doordash giftcards! and would be really appreciative. DM if you're interested please.

[kindred garnet]

do u have any neg prompts for avoiding double belly buttons? or some weird lines on the belly?

[brisk badger]

Hello, i have this photo edited fast in GIMP, i would like to do it more Realistic, like they're really fighting, all prompts like "one white man fights one black man and one white man with guns, 8k, realistic," only give some vague characters, thanks for your help

[lone charm]

Have you tried getting a decent result then inpainting that? Usually the best way in my experience, usually I get like one good out of 5 so gotta do a decently sized batch

[wanton marsh]

Hello guys, kinda new here 🙂
Do you guys now what prompts i have to use to make an specific character?
The character is Dehya from Genshin Impact with the sky on Fire
(sorry if my english is not good at all)

[smoky totem]

short of using a lora, you just just try to use that specific description Dehya from Genshin Impact, background sky on fire, as starting points. I bet there are many images of Dehya in the database already since genshin is recent and well known.

[wanton marsh]

thanks buddy, any recommendations about where i can find Lora models?

[smoky totem]

civitai

[fresh yacht]

Can someone enlighten me to as what {tag} does compared to (tag) and what doing (tag:1.5) do?

[pliant rapids]

What prompts would you recommend to generate a picture of a bengal cat (the house pet)? I keep getting bengal tigers, or cats that look like bengal tigers. Left is what I'm going for, right is what I get. Using the 2.1 settings on the huggingface website.

[smoky totem]

tried negative prompt tiger? also positive prompt bengalcat and bengal_cat

[slim orbit]

How is DDfusion different in SD than Google colab

[analog latch]

Try "domesticated Bengel cat" or "Bengel Longhair" or "Cashmere Bengal"

[lone badge]

wow ! nice and well done ! I didn't manage to make that one work

[sour scarab]

Is there a way to assign different colors to a subject and a background? If I write "purple background" the subject often also turns purple and the other way around

[crimson bay]

anything diffusion 4.5 just gives me dull images and faint colors, only with cherkpoint or Loras to get more vivid color results?

[silver valley]

you need the .vae.pt file from anything v4

copy and rename it to match v4.5

[crimson bay]

I am using colab, can you tell me in which folder on google drive I put it?

[silver valley]

dont know how the folder structure looks like, but normaly it goes inside the folder where the models are stored:
Rename the vae to match the model like:
Example.safetensor
Example.vae.pt

[crimson bay]

thank you very much

[silver valley]

no problem 🙂

[kindred garnet]

Do u guys have any ideas to get the same or similar style like in midjourney ? i tried several models / prompts but it doesnt really bring that midjourney look , it looks more painted/ soft in midjourney

as an example of a midjourney pic

[barren turret]

did you try openjourney?

I guess they are trained on the data of midjourney

[kindred garnet]

yes, but feel like the faces are more versatile, unique and it feels more like "art" less photorealistic

[barren turret]

ah

[kindred garnet]

just an overall different look

[barren turret]

I'll try a few prompts w it if i can get something similar then would let you know the prompt

[kindred garnet]

oh nice ty 🙂

[atomic flume]

how do I get SD to put the eyes of an image into sharp focus ?

Negative prompt: (deformed iris, deformed pupils, semi-realistic, cgi, 3d, render, sketch, cartoon, drawing, anime:1.4), text, close up, cropped, out of frame, worst quality, low quality, jpeg artifacts, ugly, duplicate, morbid, mutilated, extra fingers, mutated hands, poorly drawn hands, poorly drawn face, mutation, deformed, blurry, dehydrated, bad anatomy, bad proportions, extra limbs, cloned face, disfigured, gross proportions, malformed limbs, missing arms, missing legs, extra arms, extra legs, fused fingers, too many fingers, long neck
Steps: 50, Sampler: Euler a, CFG scale: 7, Seed: 4186429302, Size: 600x600, Model hash: c35782bad8, Model: realisticVisionV13_v13, Denoising strength: 0.5, Hires upscale: 2, Hires steps: 25, Hires upscaler: 4x_NMKD-Superscale-SP_178000_G```

[tribal spruce]

Hello. I am trying to generate an image of a burger. The burger looks great but it's always cut off on the edges. Is there a way to prevent that?

[smoky totem]

put cut off image or cut off subject in negative prompt, and/or try to use outpainting to "complete" the burger image by adding to an existing finished image.

[smoky totem]

try put focus on eyes earlier in the prompt and soft focus/out of focus earlier on neg prompt

your neg prompt right now looks like a standard one I've seen from a model maker.

try sharp focus on eyes too

[tribal spruce]

Oh my god thank you

[kindred garnet]

@tribal spruce can we see ur result? 😄

[tribal spruce]

Things like this. It's really in love with tomatoes. I wanted to take the focus off of the tomatoes and next result I got was tomatoes on the burger and ON the burger, as in on top of it

[kindred garnet]

h nice

[tribal spruce]

Now it burnt the tomatoes for some reason...

[brisk badger]

Thank you very much but How to inpaint it?

[tribal spruce]

It started cutting off again. Also it's in love with tomatoes and I don't know how to reduce it and it started building towers...
Prompt: burger with a few tomatoes, onions, cheese, salad, front view, center, subject in center
Negative prompt: cut off, cut off image, cut off subject

[tribal spruce]

This is what it produces

[silver valley]

your prompt in realistic vision:

[tribal spruce]

Medium rare

Currently it just cuts off the burgers in every single run. Sometimes the burger has 5 patties or too many tomatoes or stuff like that...

Right one even features fresh imported tomatoes right out of the heart of Tschernobyl (Ik that it's AI and things like that happen, I just think it's funny. All I want is the burger to not be cut off)

[silver valley]

which model do you use ?

try others

[tribal spruce]

It did not cut it off for a while but went crazy with tomatoes. I use Heun btw, if you mean that

[silver valley]

no i mean model

"checkpoint"

[tribal spruce]

Yeah, just realised. sd v1.4

[silver valley]

yea, i would suggest you try 1.5 or 2.1
or community made ones like: Realistic Vision or Illuminati Diffusion

[tribal spruce]

My only big issue is just the cutting. The burgers look great. I will download another one I guess but they just need around 2-3h with my internet

I mean, look at this one. It's perfect! ... but cut off

[silver valley]

yea thats kind an issue of official models, they can do everything but not everything right ^^

[tribal spruce]

This one is not cut off but it looks too clean. It has this artificial look

Or is that a sampler thing? Because I think mine looked the same in the beginning but on Heun they look great

[silver valley]

it depends on the model mostly

and the prompt

[smoky totem]

It looks like there is a straw on the burger. I learn a new way to enjoy them now.

[tribal spruce]

It just reaaaallly loves tomatoes XD

Any improvement suggestions on the prompt? Otherwise I would try to download another model but my internet is really slow.
Prompt is burger with one patty, a tomato, onions, cheese, salad, front view, center, subject in center, padding around burger <- Desperate try to prevent it from cutting
Negative prompt: cut off, cut off image, cut off subject

[lusty crow]

I fail when try to simplify the prompt.

[tribal spruce]

It even made a store😮

That's cool

[lusty crow]

And then the tomato cheeseburger with extra tomatoes.

[tribal spruce]

LMAO

Even the bun is half tomato

[tribal spruce]

Fuck it, I'm gonna include that burger into my UI project (if I`m allowed to?)

[lusty crow]

It okat 😄

okato tomato okay-to

[tribal spruce]

XD

[lusty crow]

I added "trending on artstation".

[tribal spruce]

What model is that btw?

[lusty crow]

The first (tomato) one was Deliberate and the last was V1.5

[tribal spruce]

I guess I'm gonna download v1.5 then because all of the images v1.4 generates at the moment are cut off and I desperately try to fix it in the prompt but it just adds tomatoes

[lusty crow]

"one center placed tomato cheeseburger seen from front, with (extra tomatoes:1.2)"

And V1.5 do fail from time to time and place it out of frame, it AI so it isn't smart.

And can give images like this too :/

[tribal spruce]

v1.4 fails all the time to fully place it inside. It did for a short time but now it doesn't anymore. I just generated images where it took the burger apart and it looked like salad

Now I seem to get a burger and a building

Changing the prompt was not a good idea

I just wanted to make a UI with AI generated assets and needed a burger. Now I spent half of my day generating burgers...😭

[lone badge]

are we prompting burgers ?

I did a burger model once

[tribal spruce]

I'm trying. They look good but are always cut off and idk why

[lone badge]

https://huggingface.co/Guizmus/SD_PoW_Collection#burgy

trained on 50 delicious pics of burger

[lusty crow]

You can also try to say what is on the top and bottom to try to frame the subject: "one (center placed) double cheeseburger with tomato on a napkin seen from front (extra tomato) flag on top".

[tribal spruce]

How big is the file?

[lone badge]

it's a ckpt, 2GB. sorry, it was done before LORA was a thing

[tribal spruce]

Currently downloading v1.5 hoping that it can center it better than v1.4. 2h left...

[tribal spruce]

I only have CKPTs atm

[lone badge]

there is the ckpt https://huggingface.co/Guizmus/SD_PoW_Collection/resolve/main/291122/ckpts/Burgy.ckpt
and here is the dataset if you want to check what I trained on https://huggingface.co/Guizmus/SD_PoW_Collection/resolve/main/291122/dataset_Burgy.zip

token to use in prompt is "Burgy"

lol never thought this model would be of use again

[tribal spruce]

The burgers v1.4 creates look bautiful (maybe it adds too much tomatoes) but it just cuts them off most of the time. The one I replied too would have been at least close to perfect, but it's cut off...

[lone charm]

do u use automatic1111's version?

[tribal spruce]

Downlaoded 1.5 now but it just doesn't get any better. I think I got the best result in the beginning and from then on it has just gotten worse and worse...

[rare parcel]

please how do i find the right model for making drawing sketches

[runic osprey]

#1047197565365538826, civit.ai, google, make it yourself

[rare parcel]

im a bit slow please tell me do you mean a model called google or you mean i look it up

[runic osprey]

i mean do a google search for something like "stable diffusion 2.1 sketch style embedding"

[rare parcel]

oh thanks the thing is i dont even know the keyword to search something like this. like this " style embedding" for example

[runic osprey]

there are different types of "models" (idk if that's the right word), which improve stable diffusion prompts. embedding is the simplest / smallest / easiest to install (though probably the worst results)

[willow trout]

I know brackets give words more priority, is there something similar to give words less?

[rare parcel]

thanks for the info! sadly nothing on civit.ai and a google search didnt bring anything up thats why i was asking here if someone used one

[runic osprey]

square brackets []

[rare parcel]

i tried mid journey and it made some nice sketches and i was wondering if i can do it in SD

[runic osprey]

if theres a very specific style you want, that no one else has done yet, you might have to train your own

[rare parcel]

no not really specific quite the opposite but i dont know how to train

how many images approximately do i need to train it

[runic osprey]

for embeddings it can be quite small, like 10-20

idk how much others require

[rare parcel]

oh niice

can you link me a tutorial please

[runic osprey]

https://www.reddit.com/r/promptcraft/comments/zyc5eh/stable_diffusion_detailed_tutorial_on_embeddings/

[atomic flume]

I was generating goth models and after seeing your cheese burger I took a shot at making them hold a cheese burger

[mental hound]

what are some key words I can input so that my legs are fully clothes and not showing skin? doing impaint job.
I have tried naked, nude, skin, but they dont do the trick

[feral fjord]

Maybe mention the clothing piece you want them to wear. Like trousers, jeans, long dress, etc.

Assuming you mean what to put in the prompt and that you tried naked, nude, etc. on the negative prompt

[mental hound]

thanks will try that.

[feral fjord]

Also, does anyone know what I could add to the prompt it so this picture doesn't look so pixelated? And also so the architecture is a little bit clearer instead of a visual buzz that doesn't make sense haha

The pixelation is not that big of a problem since I can fix that with resizing with ESRGAN, but the architectural nonsense is a problem I can't seem to get rid of

Even if I add words like "sharp details, extremely detailed" etc

[atomic flume]

what resolution did you set it too ?

[feral fjord]

W 768, H 576. The pixelation at least didnt show up in previous results. I must've messed up the prompt or something

My concern is the architecture. Do you think a higher res will provide better results?

[atomic flume]

A little bit more could help are you using high rez fix?

[feral fjord]

Nope

Should I?

[atomic flume]

it can raise the resoution much higher

so give it a try see if it fixes your problem

if not then we can start messing with some other settings

also how many steps are you doing

[feral fjord]

Alright, I'll give it a try

Thanks

[feral fjord]

A low number for now, between 15 to 20 with DDIM

[atomic flume]

try going up to 40

[feral fjord]

Oh dang. I'll try that and the high res

Thank you for the tips

[noble sigil]

I've always written my prompts like: "medieval, knight, heavy armor" but lately I've seen people write something along the lines of "medieval knight in heavy armor swinging huge sword" is there a difference how SD interprets both prompts?

[feral fjord]

Hey, I left it for a while and I tried again and your tips actually improved the image by a lot. I used the same settings by using PNG info and then I did 40 steps with DPM++ SDE and applied hires fix. It looks way better now and after further upscaling I think I'll use this one. Thanks a lot!

[fresh nest]

@silver valley

I am trying to run a list of prompts on a custom model I made and over several ckpt’s saved on different steps for that training session.

Each ckpt has the same token: shbdg

[silver valley]

Is shbdg specific to the model ?

A style or something ?

[fresh nest]

Yes.

Trigger token

Dreambooth model

[silver valley]

Ah okay

Your Problem is you need a replaxer word before your tags

Replacer

The word dont need to exist.
Type for example
lolol at the start of your prompt.
Then add lolol, as first word in x values

[fresh nest]

Only add it one time but first?

[silver valley]

This word will then be replaced with the followed words

[fresh nest]

Ok. But there is no issue that shbdg is repeated in every prompt?

I thought that was the issue

[silver valley]

No that shouldnt be an problem

[fresh nest]

Ok. I will try rn

Thanks!

[silver valley]

You can also hover over S/R Prompt to get more Information

[fresh nest]

Aah! Didn’t know that 😀 thanks x2

[silver valley]

Yea its not known but most stuff is good documented

[fresh nest]

But should I also add that word first to the regular prompt?

[silver valley]

Yes like i said

Prompt S/R stands for Search and Replace

So it will look for that word to replace it with your stuff

[fresh nest]

But I mean: put it both in the regular prompt and in X values where I keep all following prompts?

[silver valley]

Yes it searches for it in your prompt

It needs a match

Then it will replace it with your stuff

[fresh nest]

Awesome. Now it works 🙂

[lone badge]

Hey Joachim 🙂
In your case, you could have a prompt like

shbdg a woman wearing a brown jacket in a city closeup portrait shot bokeh photo
then have params like those ones for example, to test 9 different combinations of prompts on your checkpoints

(sorry was very very slow :p but CS1o rocks)

[fresh nest]

I don’t understand.

[silver valley]

Yea thats next level S/R

[lone badge]

the prompt has 2 words that will be researched in this case : "woman" (that will be replaced by man and child), and "brown" (that will be replaced by red and purple), so 9 total prompt

and the lot is run once per checkpoint

giving a grid for each checkpoint

[fresh nest]

Another thing: if I want to use commas in my prompt, should I then put “” around them?

[lone charm]

🤔

[lone badge]

not sure I see any trick for using commas in S/R

why would you replace the comma though

[fresh nest]

I don’t want to replace commas

[lone badge]

you aren't supposed to replace the whole prompt, just some words to make it cycle through some tokens ^^

I get what you are trying to do though

[fresh nest]

I want to use prompts with commas in them without confusing it with the comma that separates the prompts.

[lone charm]

i think they want to exchange a string of words like "cat, photo" with "dog, drawing" for example, but ye not supposed to replace a large section, usually single words

[fresh nest]

What do you mean “supposed to”? I want to try out some basic prompts throughout my different step ckpt’s from a training season. What’s wrong with that? 🙂

[lone badge]

there is another script that does something maybe more fitting, "stresstest a list of checkpoint on multiple prompts", never used it yet though.

[lone charm]

o ye u can put the prompts in a text file then run a prompt batch i think

[lone badge]

the "prompt S/R" is "supposed" to, I meant the feature has been thought for, replacing parts of a prompt, not the full prompt. if going for a list of prompt, other tools won't have problems like the comma thing

but yeah, outside of the comma thing, your way should work too

[silver valley]

I know i read something with \ for adding stuff like () or ,

Cant find it

[fresh nest]

I will have to get into this on another occasion. 🙂

[silver valley]

Found it, discord breaks the Syntax when i copy the Text

[fresh nest]

Ok. So I will have to do it like this in prompt then \,\

[silver valley]

Yea maybe that works

[fresh nest]

Ok

[silver valley]

Never tried it

[fresh nest]

Ah ok

[still grove]

How do I get the basics to work without Models-Embedding and all that jazz?
I do admit I am new to all this stuff, I for instance don't even know if I am working with 2.1 or 1.5, I think 1.5, because of the SD15New in the top
The thing I am trying to achieve is somewhat correct proportions, as in, a classing dungeons and dragons dragon in this case.
There are a few things I suspect, like putting a list of dont do ugly things in negative, or maybe it has some difficulty because it is less experienced with drawing dragons than say humans, however what is the best way to get started to achieve more consistent results?

just adding modifies apparently really helps, still not fully satisfied with the silly looking head and fore limbs, but a lot better

[lone charm]

heres a negative prompt i found that usually works well, usually for people tho:

blurry, rendering, photography, painting, signature, (ugly), (duplicate), (morbid), (mutilated), (mutated), (deformed), (disfigured), (extra limbs), (malformed limbs), (missing arms), (missing legs), (extra arms), (extra legs), (fused fingers), (too many fingers), long neck, low quality, worst quality,(Wireframe),Polygons,Screenshot,Character design,Software,UI,(watermark),(text),(overlay),getty images,(cropped),low quality,worst quality

and a very basic prompt i usually start with, just start adding stuff:

photorealistic, highly detailed, beautiful, 4k, 8k, trending, award-winning

[still grove]

also, is there a way to combine elements,
I started off with red dragon on a hoard of gold,
It somewhat resembled a gold dragon with a red background,
Is that just a fluke, if not, is there a way to make it do something like that?

[lone charm]

try running it in a batch of like 4, usually i get 1/4 good ones

[still grove]

I dont know if this is the right place, however:
One of my goals with stable diffusion is to easily make dungeons and dragons character portraits, is there a good model that specialises in drawing fantasy characters (as in, wizards, necromancers, knights that kind of thing), preferably in a semi realistic or not anime style?

[silver valley]

Try Dreamshaper or Protogen

[mental hound]

Does it matter if you use (by x artist:1.3), (by y artist1:3) or by x artist, y artist

So does (:1.3) do anything?

[mental hound]

Try RPG4

[feral fjord]

Yes. There's a few. Depends on if you're more into illustration/drawing type or if you're into realistic portraits

RPG4 is freaking good. But! It's kinda complex to use, there's a guide that fully explains how. You can find it here https://civitai.com/models/1116/rpg

And for more illustration/painting type of art I use Dreamlike Diffusion (https://huggingface.co/dreamlike-art/dreamlike-diffusion-1.0) or Dreamshaper (https://civitai.com/models/4384/dreamshaper)

[mental hound]

I made this with RPG4. The one Fran suggested

[feral fjord]

Amazing

[mental hound]

Thanks

[still grove]

nice, ill look into it

[sour scarab]

Having some trouble getting an earth genasi rendered (stone or cracked stone as skin)
positive: face portrait earth genasi, gray cracked dirt and scales as skin, cosmic background, yellow wizard robes, black smoke hair, photo, kiss, 80s
neg: woman, cropped, lowres, poorly drawn face, out of frame, poorly drawn hands, blurry, bad art, blurred, text, watermark, disfigured, deformed, closed eyes

I'm getting stuff like this which is awesome of course, but not exactly what I'm looking for

Maybe I can get cracked skin in with inpainting, but I haven't had much success with that either. I got closer with stable diffusion 1.5

[hexed sandal]

Hello. I`m trying to figure out why I dont get the same pictures in SD when following the exact same promts and seeds model etc from Civitai. Could anyone please explain ?

[tired vigil]

Are you using xformers?

[bleak fractal]

So I’m trying to generate swords but no matter what it never generates a good looking sword, I’m guessing stable diffusion isn’t trained on a lot of items? This issue happens on a lot of items. Staff, pickaxe, pen, wand, etc.

[smoky totem]

we talked about this a couple of days ago, and earlier than that. search and read up on it. not discouraging you asking, just do not want to explain again.

[honest wren]

guys im having this issue, i get greyish outputs instead of what im supposed to get copying prompts from internet

that is an example, do some1 know the reason why this is happening? Ty in advance

[smoky totem]

@tranquil folio hi, what is the trigger word for your contrast fix lora(s) for 1.5 and 2.1? When I look at your examples and others from the community on civitai, I did not see triggers like lora:theovercomer8sContrastFix_sd21768:1. Is it automatic? I tried to use it with a 2.0 model (dont know if that was a problem) on a bright day scene, and its effect seem minimal if at all. I tried to bump up to :1.5 as well.

I put lora:theovercomer8sContrastFix_sd21768:1 on the positive prompt section at the end.

the 2.0 model does robots but trained at 512x512 so that maybe a problem?

I can't really tell if contrast fix is applied. I had to use a white mech to test the shadow and contrast since dark ones come pretty contrasted by standard. However, with noise-offset influenced models, I can see distinct vignette on the images, and this one does not have it. As such I am assuming that your lora is not triggered.

[tranquil folio]

to8contrast style is the trigger word i listed on civit. i generated the samples using the additional networks plugin which doesn't put the lora into the prompt. it should have the same effect by putting the lora in the prompt and the trigger word

[smoky totem]

thanks, have you tried it on 2.0 models? are 2.1 trained loras backwards compatible with 2.0 (not 1.5)?

[tranquil folio]

i haven't tried it on 2.0

i'd guess it'd only work on 2.1

was there a 768 for 2.0?

[tired vigil]

You are probably missing vae,

Edit:
read through model description and files included and see if you're missing something

[smoky totem]

no sorry that robo model is 2.0 trained at 512x512

[tranquil folio]

its a 768 lora

dunno what it'll do on 512

[smoky totem]

will the 1.5 lora work on 2.0 (both 512)

I'll just try it and see how it goes

[tranquil folio]

i doubt it 😦

[smoky totem]

ok thanks anyways I will find a solution. I will make my own lora based on the robo checkpoint images I generate to "upgrade" it to 2.1 (since it can generate 768x768 images with no problem), then I will use your contrastfix for 2.1

[smoky totem]

I'll try dvmech as an alternative, its style is a lil different from the 2.0 robo one, but it is trained natively on 2.1 at 768x768, should play well with your contrastfix lora

[next wharf]

So I’m trying to make some beautiful oc character sheets for personal use, but the Ai is giving me weird looking people?

[past sluice]

has anyone worked out how to make SD2.1 reliably render a couple where the woman is taller?

I totally get that it's only going to find nearest paths based on ingested data so it's not trying to be sexist, but I've been tinkering for about 20 mins now trying to make it produce this, and the closest I've been able to get is a woman dramatically in the foreground and the male partner in the distant background lol

annoyingly, even when I feed it img2img with a bunch of different illustrations in the correct pose I'm trying to produce (a gender reversal of the famous ww2 "soldier kissing nurse on may day" portrait) it forces the existing female character to become male and vice versa, even on super low denoising strengths

interestingly, I can kind of get it to produce platonic "standing side by side" images, but the second I introduce the phrase "kissing" it spikes hard in the way described above

would love any ideas on how to trick it out of it's gender stereotypes lol

[atomic flume]

Hey I'm trying to create something like this in SD

any idea how I can do that I wanted the female character to have golden finger tips like this with the gold all liquid

then another image more like this

Where the fingers are just gold

[feral fjord]

Perhaps inpainting or photoshop? Generators aren't usually good with very precise details

What you could do is inpaint the hands, use multi ControlNet with Depth and Scribble to ensure that the hands are in the exact position you want and prompt it to make the new hands with golden fingers

The alternative (which is wayyy simpler) is just using photoshop

I'm sure you can find a tutorial to make this effect

[dusky aspen]

Hey everyone

How do I zoom out and give a character new body parts

[lone badge]

that's called "outpainting". there is a script for it at the bottom of the img2img tab

also, the other tool outthere, invokeAI, is really good for this kind of things, let me find an example

https://www.youtube.com/watch?v=tV9aQtUzt34

[silver valley]

Outpainting with the openOutpaint extension or the outpaint script, but you need a special model for it (You can also convert the others to use)

[tired vigil]

What prompt would generate identical faces as per the source image?

[dusky aspen]

I MADE FEET WITHOUT OUTPAINT \o/

THEY'RE SO CUTE

[feral fjord]

Damn, they're pretty decent

Nice job!

[feral fjord]

It literally looks like its made by a person, there's almost no AI artifacts

Congrats!

[fresh nest]

what's your best negative prompt to help with bad anime eyes?

[humble sedge]

no matter which model I use I keep getting anime and can't prompt my way out of it

[silver valley]

Try realistic vision or deliberate model

[humble sedge]

yeah deliberate is usually my go to and sometimes the results are great, but keep getting the same characteristics as anime e.g. big head, eyes, pointy chin

its almost like the closer the picture is, the more anime it becomes, the more zoomed out, the more f222 it becomes

e.g these were in the same batch

[trim ether]

anyone know what prompt for artstyle like this? and maybe the model too

[past sluice]

This was a marathon but finally got there by starting with two women (which was the closest it would let me get with txt2img), then inpainting the male character slowly in parts 😮‍💨

[rare gull]

Hello! Anyone know a prompt/model for this sorta rough hand drawn artstyle? I'm in love with it.

[lone badge]

did you try to "Interrogate" CLIP to see if you find some artists name that would fit ?

[rare gull]

Yes I did, and unfortunately the results weren't too great. I'm not sure if it's because of my model, but it mostly resulted in these sorts of images.

[upbeat smelt]

Hello! I am very new to Stable Diffusion, in fact just learned about it earlier today and managed to install sdwebui and played around with different models. I'm very interested in copying screenshots of video game characters (Lost Ark right now) and upscaling them to look realistic and very detailed. I attached 2 pictures of someone basically achieving what I'm looking for, staying extremely loyal to the source material, which I don't seem to be able to find out how to do (left is in game, right is the output). Any tips? :D

I know the model used is anything-v4.5-pruned.safetensors with anything-v4.0.vae.pt, other than that, nothing really hah

[lone badge]

hum.... maybe yeah... I'll try to see what I can get to

what I got through all pics as useful style keywords :

, Dan Content, official art, a storybook illustration, sots art
, Clara Miller Burd, sepia, a character portrait, synthetism
, Caroline Chariot-Dayez
, a detailed drawing, mail art

I can't get anything coherent either....

sorry, I'll try again later, but yeah, a model could help, I'll look on civitai

[lone badge]

"upgrading" the art for more details and upscaling will be 2 different things, maybe try to focus on the details first. The techniques used here seem multiple :
img2img to keep the colors and initial composition
controlnet with canny mode (not sure on the mode) to keep the shape close to the source and not stray too far in the changes that happen
a good prompt to help push the AI into adding the good details
the good model to fit with this fantasy realistic style
potentially embeddings and/or hypernetwork to push the quality of results even further

[silver valley]

its latent upscaled, so you can also upscale it with img2img sd upscale script

[feral fjord]

Does anyone know of any models that are focused on generating objects/items? I'm trying to make images for an RPG campaign. When I try to generate a wooden staff it generates this kind of images

Or this

I'm using dreamlike diffusion

Which is cool and all but it's sort of useless for what I'm tryna make

[upbeat smelt]

Gotcha! Installed ControlNet and playing around with settings, which finally stays true to the contour of the original! Will play around with stuff to see how to keep the color scheme perfectly the same and only add detail, then when I have a good output, I'll stray into upscaling, as @silver valley suggested, thanks :D

[next mulch]

Any suggestions on how to get blending between two nouns? For instance if I just use "giraffe daffodil" or expand on that with '"giraffe made of daffodils" or "giraffe wearing daffodils" I still tend to just always end up with either a giraffe, a field of daffodils, or both in the same scene. On a rare occasion, some attempts produce some flowers replacing the tufts of hair on the top of the giraffes head, but that's about the extent of it. This is just one example, but I've found that whenever I use very distinct nouns, I never have much luck getting them to blend together somehow.

[tired vigil]

how do i deal with the top of the head getting cut off EVERY time

[silver valley]

try cowboyshot or portrait

[tired vigil]

doesnt help much, just changes the style 🥀

[silver valley]

which model? its also missing the vae file for it

[tired vigil]

vae?

and the model is anythingAndEverything

[silver valley]

a vae is used and needed for anime models for color correction, i can you show you an example shortly

what tags are you using and what resolution?

try a resolution of 512x768 and then use for example:
portrait of a girl ....

[tired vigil]

res is 512x768

[silver valley]

oh ok

@tired vigil here is the vae difference:

[tired vigil]

oh

wow

[silver valley]

so you need to get the vae of AnythingV3 for your model

from here for example:
https://huggingface.co/AdamOswald1/Anything-Preservation/blob/acc67d36406e41252aa936c43248c4ad988db33f/Anything-V3.0.vae.pt

it goes into the models folder, then you have to rename it to match the name of the AnythingAndEverything Model.
Example123.safetensor
Example123.vae.pt

---

for your cropped problem, try to describe her hair more

like straight hair, streaked hair,

[tired vigil]

and i assume you have to apply it in the settings?

[silver valley]

no if the setting for vae is on Automatic, it will automaticly apply matching vaes to models with the same name

[tired vigil]

oh

oh wow that looks way more vibrant

also is there a way to get rid of more annoying small details that refuse to go away, such as this on the right arm

[silver valley]

you can try img2img to get similar pictures or you can inpaint and maybe remove it

[tired vigil]

@silver valley i used outpainting and after (quite a bit) of trial and error i got something that worked well

[silver valley]

nice 😄 👍

but why outpainting and not inpainting ?

[tired vigil]

for the head being cutoff, i mean

i'm using inpainting rn to alter some other stuff

[silver valley]

ahh yea good idea with outpainting then

[tired vigil]

ended up with this

[silver valley]

wow very good 🙂 the outpaint and inpaint worked well

also much better colors 😄

[tired vigil]

yeah

the outpaint was an absolute nightmare though

[south solar]

When you are using prompt editing with A1111, does SD take into account the how the prompt will change before it starts to generate the image? I am adding something halfway through generation, which should be late enough that it doesn't affect the composition, but depending on what I try to add, it totally changes the whole scene.

It is almost as though the presence of the elements to be added affect the weighting of other elements even before they start to be drawn.