📝｜prompting-help

[silver valley]

Np, most people use 512x768 for portrait and then upscale

[leaden socket]

Still do 512x512 then upscale 4x on width?

Got it, thank you!

Could I ask why 768 and not 1024?

Is that because 768 is what models are trained on?

[silver valley]

You can also do 540x960 upscale by 2 to get FullHD

[leaden socket]

Would you suggest 960 or 768?

[silver valley]

For 2.1 models you could go 768x1024

[silver valley]

768

But you can play around. Its not a must

[leaden socket]

I assume this is the version?

Thank you, this is all new to me.

[silver valley]

No problem 🙂

Whats your gpu btw?

And do you know how to upscale ?

[leaden socket]

You think this is good?

    parser.add_argument("--width", type=int, required=False, default=512, help="Width of the output image")
    parser.add_argument("--height", type=int, required=False, default=512, help="Height of the output image")
    parser.add_argument("--hr_resize_x", type=int, required=False, default=2048, help="Width of the output image")
    parser.add_argument("--hr_resize_y", type=int, required=False, default=1024, help="Height of the output image")

[silver valley]

Ohh you use the Commandline Version 😮

[leaden socket]

yea

API

[silver valley]

Ahh okay

Idk if that will work this way

[leaden socket]

Testing.

[silver valley]

Why not using the webui ? ^^

[leaden socket]

Easier to do more batching / automate.

[silver valley]

Ah okay

[leaden socket]

It works, I just have to disable NAN checks. Thanks for your help!

No idea how but I broke my automatic1111:

modules.devices.NansException: A tensor with all NaNs was produced in Unet. This could be either because there's not enough precision to represent the picture, or because your video card does not support half type. Try setting the "Upcast cross attention layer to float32" option in Settings > Stable Diffusion or using the --no-half commandline argument to fix this. Use --disable-nan-check commandline argument to disable this check.

a reset and adding those flags didn't fix my issue.

[leaden socket]

One final question for you, would you recommend any changes?

[oak nova]

That upscale is a different aspect ratio.
Unless you have a 4090, that upscale might not be possible (too high res).
If you want an upscale like that, use the SD Upscale script in Img2Img.

[leaden socket]

I can upscale it.

[leaden socket]

Anyone know why I would be getting some pretty bad upscaling?

    parser.add_argument("--hr_scale", type=int, default=2, help="High resolution scale factor")
    parser.add_argument("--hr_upscaler", type=str, default="Latent", help="High resolution scale factor")
    parser.add_argument("--hr_resize_y", type=int, required=False, default=2048, help="Height of the output image")
    parser.add_argument("--hr_resize_x", type=int, required=False, default=1024, help="Width of the output image")

[silver valley]

Yes, first choose 10-20 hires steps, thats enough.
Then set the denois value to 0.5 or less because you select an esrgan based upscaler.
Then i would suggest to use the "Upscale by" setting instead of setting a resolution.

[leaden socket]

use the upscale by Thank you, if I want portrates how would I do that if I do 512x512

[silver valley]

You would use 512x768 or 540x960

Upscale by 2 would get you:
1024x1536
or 1080x1920

[tired vigil]

/help

[lost crest]

How do people get images where a woman is like half skeleton, half human

[rich crow]

Could be inpainting or could be regional prompting or could just be a really good cherry picked seed from a good prompt

Editing exists too

[steep fossil]

Where are some pages with info on good prompt addition for styles, artists' names we can use, etc? I'm trying to make a comic and just appending "illustration, comic" .. and all my generations look almost exactly the same style

which imo is kinda weird itself.. I'd imagine there to be a lot more variation in what an illustration, comic looks like.. but ok

[thin plume]

Soo. I'm using Automatic1111's webui, and I'm using the [ Alternate | between ] things command. So if you do [Dog | cat] it should generate something that's half dog, half cat, or whatnot, right?

Anyway. I've been blending together some celebrity faces. It's been interesting, and I'm still learning! Although, I've just ran into a weird issue. When it's generating the image, things are looking fine and good! As soon as it finishes though, the end result looks.. Well, nothing like what was being generated (in terms of the face.)

Can anyone enlighten me as to what's going on? Is there a way I can just have it stop at the 2nd to last iteration, so I can get a final image with that face, instead of the one it's giving me?
This isn't a one-off thing btw! I've had this issue a handful of times now.

[vapid moat]

Do you happen to have face restore enabled?

[thin plume]

I do! Is that the issue?

[vapid moat]

Yes! Try disabling it

With newer models face restore usually makes things worse

[thin plume]

Ehh, nope. That made it worse

[leaden socket]

Anyone know what's the best negative propmpt for text?

text, logo, caption, description, font, words, text not working

[vapid moat]

watermarked?

[steep otter]

Suggestions?

[silver valley]

Looks nice, you could try upscaling with hires fix for more Details.

[steep otter]

Doing that right now... takes quite a bit.,,

[silver valley]

Whats your gpu ?

Try upscale by 2, denois 0.5, hires steps 10, upscaler: esrgan4x+ anime6b

[steep otter]

Oh, I thought you want me to sample again...

I started with a 512x512, did a latent upscale to 768x768, then sampled it again.

[steep otter]

GTX1060...

[silver valley]

You can do 512x512 and upscale it by 2 with highres fix, no need for resampling

[silver valley]

With 6gb or 3gb vram ?

Make sure you use --xformers --medvram to get the most out of it

[lost crest]

Is it with inpainting the intention that I keep the entire prompt and only change some words or do I wipe the entire prompt and use the words to describe what needs to change in the area that I masked?

[steep otter]

6 GB

[steep otter]

I’m using xformers, medram isn’t necessary

[steep otter]

Wait… isn’t hires fix fundamentally just resembling tho

[wintry falcon]

Where to put the promtes in the stable diffisuion to generate images

[silver valley]

No its an option you can active in txt2img

[steep otter]

But isn’t that option fundamentally doing a latent upscale then resampling?

I’m not using AUTOMATIC1111

[silver valley]

Ahhh okay

What do you use ?

[steep otter]

https://github.com/comfyanonymous/ComfyUI

[silver valley]

Ah comfyui, i dont know how its done there but there is a seperate tool for upscaling similar to comfy called chainner.
https://github.com/chaiNNer-org/chaiNNer

[steep otter]

Oh, thanks

[solar carbon]

when I type a prompt like "a beautiful woman wearing a hard hat and painting a fence" I get duplicates. I almost always get duplicates when trying to show one person. How do I fix this?

[obtuse torrent]

@solar carbon, the standard tip, for A1111, start at low res (960x512) + check Hires.fix and use "Upscale by" (x 2 is a good factor, ) , add "solo" to positive prompt, and add in the negative prompt: (cloned), (cloned face), , (duplicate), (jpeg artifacts), If you still have the problem, you can increase the weights for those tags, and/or reduce the initial resolution a little more to 920x512.

[solar carbon]

Ok thanks for the tips. Is 960x512 a 16:9?

[obtuse torrent]

@solar carbon ..for it you need 960x540 or 912x512 -- 864x486

[solar carbon]

Thx

[solar carbon]

Any good free video upscalers? I found one for framerate upscale but any resolution upscaler for video im finding wants me to pay

[obtuse torrent]

... https://discord.com/channels/1002292111942635562/1072238304042438758 re ask here..@solar carbon

[tired vigil]

do you guys know good models to create stuff like this, horror cosmic stuff

like hp lovecraft

[obtuse torrent]

@tired vigil maybe some1 can work for you https://civitai.com/tag/lovecraft

[stone igloo]

FAQ: Where can I access the DreamStudio beta website?

https://beta.dreamstudio.ai

[leaden socket]

Anyone know how I would remove this text that is always showing up in the bottom corners.

Here's my negative:
blurry, depth of field, vfx, duplicate, cartoon, sketch, b&w, black and white, text, logo, caption, description, font, words, text fuzzy, specks, floating, bad quality, pixelated, text, pornographic, erotic, eroge, hentai, ecchi, futanari, shota, shotacon loli, lolicon, bad-artist, bad-hands-5, bad_quality, (bad_prompt_version2:0.7), claws, umbrella, lowres, bad-artist, bad-hands-5,claws, fullbody, zoom out, bad_quality, nsfw,lowres, bad anatomy, bad hands, text, missing fingers, extra digit, fewer digits, cropped, worst quality, low quality, normal quality, <bad_prompt>, <bad-artist>, signature, watermark, username, blurry, artist name, abstract, picasso, amputee, deformed, blurry, disfigured, poorly drawn face, mutation, mutated, extra limb, ugly, poorly drawn hands, extra leg, bar censoring, blurry face, simple background, (multiple views:1.4), watermark, signature

[silver valley]

What are your positives?

Make sure dont use artstation, CG wallpaper or other sites for wallpapers

[leaden socket]

Basic, night scene on alien planet

[silver valley]

Ah okay, you can try add masterpiece, wallpaper, absurdres

[leaden socket]

Where do you get all this knowledge? 😄

[silver valley]

Using SD since the start last year 🙂 playing around and testing out the new stuff over time will give you a lot of knowledge. Also the wiki of auto1111 is a great source for the WebUI functions.
And of course chatting with other SD users, reading guides or watching tutorials.

[leaden socket]

I got pulled on the MJ train 😢

Waste of time.

[silver valley]

Maybe, but you tested the basics and got into SD with it

Here is also a big Source of Knowledge:
Compendium of links updated frequently about everything Stable Diffusion related:
https://www.sdcompendium.com/

[leaden socket]

I used to run Anaconda and have to deal with all the module installs for my own HF transforms, it was too difficult a year ago, automatic1111 is something else.

[leaden socket]

Thanks so much!

[silver valley]

Yea its perfect, easy to use and not to hard to setup

And we have the ultimate freedom of creating + the awesome Community expanding the SD functions

[lapis warren]

is (a,b:1.5) the same as (a:1.5), (b:1.5)?
is (a:1.5), b, c, d, e, (f:1.5) the same as (a,f:1.5), b, c, d, e?

[orchid saffron]

Is there a way to add watermark in every picture I create? I know I can use photoshop or any other editor for that but if SD can do it then I wont need other apps

[obtuse torrent]

generally, yes and yes (for tag based models)

[solar carbon]

I notice in some AI vids on youtube made with gen2, they are able to make the peoples mouths move as if they are talking then put audio over it. Is this possible with deforum on SD? When I use a prompt like "store employee talking" it just showes them with a closed mouth

[half nova]

do the vids in question have a d-id logo in the bottom

[echo umbra]

Hello, does anyone know what keywords max neural used in his prompt to generate such unique creatures for his viral astral jump video?

[wet orchid]

Is there a feature available that describes an input image back to me? So i can use it in my prompts?

[silver valley]

Yes thats called clip interrogation

Available in the auto1111 webui or trough different websites

[wet orchid]

Thanks just the info I needed!

[wheat harbor]

Umm im using regional prompter to get 3 characters
With horizontal 1,1,1 and use base prompt for the environment

The prompt layout is like this
Environment ADDBASE
GirlA prompts BREAK
GirlB prompts BREAK
GirlC
It could make 3 character but most of the time i only got 2 character with GirlB and GirlC getting combined. How can i improve the prompting for this regional prompter?

[lost crest]

I am trying to in paint the hand, however, I keep getting a result similar to this. Anyone know how I can possibly improve this?
my prompt is: yml (masterpiece:1.2), (best quality:1.2), (ultra realistic), (ultra photorealistic:1.3), (ultra-high res), (real picture, intricate details), one hand

This is not 100% the original img of the woman (I removed the nsfw part)

this is how I masked

[dusky sail]

The explanation for regional prompter is pretty terrible imo. I think you are supposed to use only BREAK in prompts. And you must put exactly the amount of BREAKs required or you'll get a broken prompt.
So for your case it would be:

base_prompt
BREAK
girlA
BREAK
girlB
BREAK
girlC

They way you wrote it you effectively merged base with girlA and omitted girlC

[wheat harbor]

Ahh okk i'll try thank you! But when to use the template it generates? It generated template as
ADDBASE
ADDCOL
ADDCOL
Which is 2D regional prompt which is very confusing if they are different than BREAK? XD

[dusky sail]

I'm not sure tbh. I just ignore the template 😆

Input the ratio e.g. 3,1,2;1, click visualize and just go by the numbering displayed

[orchid ore]

@lost crest i would go only with masked content and only masked and describe hand in prompt. In negative extra fingers etc.

[dusky sail]

Is it normal to sometimes get an "unstable" result? Or something like 2 concepts oscillating? Like for example I ask for absurdly long hair (to the ground) and on the live preview I can see it generating the rough shape and color that could be hair. But then at later steps the model forgets it was supposed to be hair and puts a random pile of cloth on the floor

[leaden socket]

This is going to be a complicated question to ask. But is there a way to have a scene and be able to change things like day, night, snow, while keeping a pretty close match to the orignal source?

and a follow up, along with changing large themes like that can you also add objects into the scene, while again keeping most of the source?

[long relic]

Ill be honest, the only good way I have found to redraw a hand or specifically a foot, is to put a reference image of what you want over the generated are. SD, when using the original image for its inpainting source (as opposed to latent noise or latent nothing) really loves to try and just make the same thing in the same place. you can set the denoiseing strength really high, like 9.5, but it doesnt mean alot except random chance.

I like to find feet that are similar enough, bend them to fit in photoshop, then inpaint it with a 0.4 or less denoise.

adding poses when inpainting the had can help. Grabbing, pointing, high five. Negatives work great too, I was trying to do a high five and adding fingernails to the negative prompt helped out.

[next ermine]

Hey how can I a throw a photo and make changes to it?

[lost crest]

use inpainting

[lapis warren]

what terms does SD understand the most / have the most data for camera angles? i.e low angle / eagles eye view, etc

the best ive been able to get is
"top view" for a view from above (but usually also tips the subject over, i.e laying on a bed),
and
"low angle", which gives a tilted picture (not quite what low angle is)

[sweet rock]

“Face” usually gives you a closeup.

[obtuse torrent]

I think that the photorealistic models have a better understanding of photographic and cinematographic terms, I speculate that the anime models fail mostly due to flaws in the tag assignment methodologies. Personally, sometimes I get the expected result with such unusual terms as zenith view, nadir view, worm view on some models, while the vast majority don't know what I mean, others recognize camera types and adjust things like a wide angle view. or a fisheye lens.

[orchid ore]

also picture layout matter very. Portrait or landscape or square. For full body view usually only portrait for example 512x768 is working properly.

[obtuse torrent]

closeup, close closeup, they work even if you don't have subjects in the composition, (someone with a face), you can use "focus" to guide the composition. @lapis warren

[gloomy kelp]

does anybody have any advice for eyes? the way im trying it now feels like luck whether theyre actually eyes or not

[silver valley]

Thats doable with the pix2pix model in img2img

[static vector]

gm anyone up? i need some prompting help

[lavish furnace]

might as well post it, im sitting here wondering if i can help but not sure

[orchid ore]

yes spit it out 🙂 @static vector

[static vector]

sorry i got carried away and forgot i asked @orchid ore @lavish furnace i am trying to output a golf cart with six seats and has solar panels on rooftop [side angled top-view of a perfect award-winning solar powered ((six seater)) golf cart with perfect six car seats with perfect space between them:2, realistic, god rays, cinematic, 8k, (symmetrical), correct car geometry, correct seating arrangement, ((correct leg space between seats))] i tried changing many words, but it always comes out with 4 or 2 seats and when it outputted the 6 seats they were crammed into one another

this is the best thing i got lol

[orchid ore]

i think solar panels can be inpainted. But they looks very nice.

try 6seater

[lavish furnace]

its beautiful
i see what he means though
no leg spacing

[orchid ore]

also do not exceed 768px for image

[lavish furnace]

silly diffusion

[static vector]

yeah people will have to put their legs out of the car lol

but how to increase the leg space

i tried everything lol

[lavish furnace]

maybe keep trying with different random seeds
or
make a quick edit yourself and do img2img

Not sure if thats the best way but thsts what id be doing 😅

[orchid ore]

yes inpainting.
4 seater but check there is space, but very small space. So they have to sit sideway 😄 . Some things are difficult to do. Or you can try if you have prompt in different model

[static vector]

i'll keep trying❤️ so i dont focus too much on changing the keywords, use same seed and try different models and inpaint

[orchid ore]

4 seater is quite common with proper space... for 6 seaters probably must be too much patience.

[orchid saffron]

What is the best resolution to use for best image generation?

[orchid ore]

should be equal to what model is trained on probably

[tired vigil]

how do i recreate anime characters?

or like any specific character because when i try to use their name as prompts it dosent really do anything

[orchid ore]

you can train Lora with them... Probably some model recognize them.

[undone acorn]

I'm trying to make pictures of my DND character who is a grung (a poison dart frog that decided to stand up) but I'm really struggling to fight the prompt bleed to get it humanoid without becoming human. If I put "anthropomorphic" it gets very sexual, even with nsfw as negative, it still makes them chiseled wearing a thong with a huge bulge... And humanoid just turns it into a person, usually with human skin too but frog legs or something. Can anyone help think of how I can make the frog be humanoid without suffering promt bleeding?

[orchid ore]

those looks like frogs at least. Nsfv filter is must.

[sweet rock]

-I'm assuming you can check metadata- parameters: ((masterpiece,highest quality)) frog, personification, animal head on a human body, knight's armor Negative prompt: ((worst quality, low quality)), NSFW, Cleavage, Pubic Hair, Nudity, Naked, Au naturel, Watermark, Text, censored, deformed, bad anatomy, disfigured, poorly drawn face, mutated, extra limb, ugly, poorly drawn hands, missing limb, floating limbs, disconnected limbs, disconnected head, malformed hands, long neck, mutated hands and fingers, bad hands, missing fingers, cropped, worst quality, low quality, mutation, poorly drawn, huge calf, bad hands, fused hand, missing hand, disappearing arms, disappearing thigh, disappearing calf, disappearing legs, missing fingers, fused fingers, abnormal eye proportion, Abnormal hands, abnormal legs, abnormal feet, abnormal fingers Steps: 20, Sampler: DPM++ 2M Karras, CFG scale: 6, Seed: 3247504133, Size: 512x512, Model hash: c0d1994c73, Model: realisticVisionV20_v20, Clip skip: 1, Version: 0acc7d3, Parser: Full parser

lol

[orchid ore]

last one from me. info in metadata 😄

[sweet rock]

Hmmm, I wonder. I looks like discord might cause metadata to get lost

[orchid ore]

i think i tried pgn info, and it worked. Will try to confirm in few minutes

it keeps it. Just dont download preview but full image 🙂

[sweet rock]

ah! Thanks for checking!

[orchid ore]

no problem

[undone acorn]

This is the funniest one imo. Thanks for the help

[sweet rock]

“X crossed with man” then describing the parts you want to be X works quite well. Thanks for that @orchid ore

[orchid ore]

yes i discovered this on my own. My advantage is that my english is very bad. I heard it doesnt make sense in english 😄

[silver valley]

Some Anime models know some characters. But at best use loras for specific chars

[tired vigil]

But how do I use/get a lora?

I downloaded one before but it didn’t recognise the file in the folder when it put it in

[silver valley]

loras go in the models/lora folder
then you press the third button below the generate button and select lora and then the lora

[tired vigil]

I put the Lora file in there but it dosent show up

[steep fossil]

I can't get a prompt to do what I want. :/ "A blue whale leaping out of the ocean with a cowboy riding it. Saddle. Beautiful Island in background."

I got a couple that almost started to get hopeful

[silver valley]

did you pressed the refresh button ?

[silver valley]

thats not a lora

[tired vigil]

wuh

o

well idk then

it said lora so i assumed it was a lora

[silver valley]

where did you find it?

[tired vigil]

whered i find what?

[silver valley]

this file

Here is a lora for example:
https://civitai.com/models/5902/aqua-konosuba-lora

[tired vigil]

i think i got the lora off of like huggingface

the file wasnt very big though so i thought something was off

[silver valley]

.bin files wont work

loras can be 2-300mb

[tired vigil]

hm

[silver valley]

ending with .safetensor

[tired vigil]

uhh

wait so can i use a lora with any type of model?

[silver valley]

exactly

[tired vigil]

i got this model called "anyloraCheckpoint" and i thought you could only use loras with that one

but if you can use it with any thats really cool

[silver valley]

the anylora checkpoint is just a model without a own style or bias so that mostly all loras would work.
But yes you can use any model. Some work better than others

[tired vigil]

ooooo

that makes sense

but what if theres no lora for the character id wanna make?

[silver valley]

you can train your own lora if you have a GPU with 6-8gb vram

[tired vigil]

i got a 3070

got like 8gb i think

[silver valley]

yes that would work

did you checked on civitai if there is the character maybe?

[tired vigil]

noo...

theres only 3 lora and none of them are of the character i wanted

[silver valley]

oh okay

[tired vigil]

is the process like super complicated to train though, some guides get really confusing

[silver valley]

you need one tool for training called Khoya_ss

installation is not that hard. but training requires some time and patience

[tired vigil]

hm

i mean if it takes time thats fine but im worried if it takes lots of thinking

[silver valley]

you can also ask some people if they would train it for you

[tired vigil]

i wan be able to do it on my own so ik how to do it for any character

[silver valley]

yea than you have to try it 🙂 but yea it involves thinking as its not that easy for the first time

[tired vigil]

you know if theres any good guides on how to do it?

[silver valley]

there was a good guide a few months ago but its outdated now

@tired vigil installation instructions are here:
https://github.com/bmaltais/kohya_ss#windows

updated the link*

[tired vigil]

uh

im so confused on where to even start

[rich nebula]

after a while of not using stable diffusion, It seems i somewhat updated it? Now if i try to use it, i get this issue and cant generate

[silver valley]

pls in #🤝｜tech-support

[rich nebula]

oh, my bad

[fleet aurora]

got a lucky shot. v2.1 is definitely better trained on leaping whales. most of them while in air showing their white belly to the viewer/camera, it would not make much sense in a rodeo situation you are looking for. that's a hard one. good luck. 🐳

[winter jackal]

Anyone open to consulting on Stable Animation UI? Trying to get results out of it

[steep fossil]

Trying 2.1. thanks!

Lol

[steep fossil]

How the heck do prompts work?

how is it processing tokens? Adjacency matters right? It's not just a list of keywords but it evaluates their relations?

(An old cowboy riding on a saddle with a seashell texture on a blue whale:1.2) The blue whale is leaping out of the ocean away from us. Beautiful Island in background. (Cowboy:1.1). Emotional western sunset. Epic photo. Award winning. Beautiful frame. Photograph.
Negative prompt: Negative prompt text: out of frame, lowres, text, error, cropped, worst quality, low quality, jpeg artifacts, ugly, duplicate, morbid, mutilated, out of frame, 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, username, watermark, signature, (cropped:1.1)

[young scarab]

Hi, does someone if/how it's possible that once you have e.g. created a prompt for a character, how you can then generate multiple different images using the same character but in different poses? I guess this is not possible w/o training you own model? (or inpainting)

[steep fossil]

@young scarab I'm not sure of a way. You might describe the character very accurately, but otherwise there's no identity detection or preservation in these things. Maybe you can reverse an image to see how the character is described, but.. no clue. Sorry. Someone more knowledgeable would likely be better to answer this q.

"(A grizzled cowboy with a weathered hat and leather chaps, firmly seated on the back of a massive, blue whale:1.3) (A full view of the massive, blue whale:1.3) (No cows: -0.7), (Cowboy not in the air: -0.7), (In the style of a classic oil painting:1.2). The whale is leaping out of the ocean away from us, with a beautiful island in the background. (Cowboy:1.1). Emotional western sunset. Epic photo. Award winning. Beautiful frame. Photograph. Yeehaw! Midrange shot." (Sand:-0.7)

Nothing's working.. grr. That was me feeding the prompt and some webpages for reference for ChatGPT to make me a good prompt.

[young scarab]

thanks @steep fossil yeah I think it would be pretty hard with just a prompt

[steep fossil]

@young scarab when this is implemented: https://arxiv.org/pdf/2304.07429.pdf

https://arxiv.org/abs/2304.07429

simple enough

[leaden socket]

Anyone know what model or prompt for something like this?

[leaden socket]

What would you guys say is a good upscaler for something like that ^

I've gotten pretty close,

[queen helm]

I'm using a DnD art model, and want to get more results like this, with 2 poses, any prompt suggestions?
Also, love when you miss a piece of the prompt that leads to wrong character traits

[sweet rock]

Your mileage may vary If anything some of the prompt on the example images might be of use. https://civitai.com/models/3036

[leaden socket]

Just look for that model in HF I'm sure there's a cell shader model.

But for the actual pose you should just prompt 2 seperate poses.

[leaden mural]

Is there a way for me to input this into sd through img2img and get it to make it higher quality?

[silver valley]

Yes you can put it in img2img and use the SD Upscale Script for example

[leaden mural]

ok ty

[covert void]

hello.

Im making training material regarding: Work safety.
The topic is Confined space.

Im trying to get StableD. to create a picture showing a chemical worker crawling down a ladder into a well.

can anybody help me with the prompt:

My current prompt.
masterpiece, high quality best quality, man, chemical technician, chemical factory, Full PPE suit, Gloves, Chemical Suit, Wearing Googles, dark blue clothe, hi-vis cloth, worker, "crawling down a well, crawling down a ladder

[gloomy kelp]

is there any way to apply prompts to a specific character? I'm trying to do two character images but saying "red hair" gives them both red hair

or just using two already established charaters

[leaden socket]

When training should I just add tags, or descriptions?

Red, helmet, visor.
Or
A character with a red visor helm

I prefer just tags. But does the nlp care?

[sonic sapphire]

Hi everyone, i'm just trying to improve my AI art. through embedding and inpainting I can correct hands and face better now, but I can't seem to do the same for feet. Are they still more difficult for AI to do well? Do I need to use a reference image? Thanks everyone

[tight stirrup]

Does anyone know if it’s possible to keep the same face rendered throughout an animation?

[sonic sapphire]

hello, I was working on inpainting and noticed that in the generation preview (2nd photo) seems to be what i'd prefer, but it turns out like the 1st image which is less detailed and darker. was wondering if anyone had advice regarding this issue. thank you

[sonic sapphire]

i tried messing with the values and it's helped a bit more

[urban parcel]

Hi I keep getting those "double eyes" almost everytime is there a way to reduce this from happening ?

[sweet rock]

are you using Restore faces? If so turn it off. It doesn't work well for anime/managa

[urban parcel]

will try that thanks

worked thanks

[teal escarp]

Boing

[leaden socket]

Can someone point me in the direction on how to train on objects?>

[sullen orbit]

I saw this prompt on civitai ''CyberPunkAI ( | neon } soldier , detailed background '' . This '' | '', is it good for something or was it just a mistake when typing the prompt?

[outer plaza]

How do I say to AI that there is a pilot inside the car? While the cars themselves are not that bad, it's being hard to convince the AI to add a pilot inside the car

[leaden socket]

It's a pipe, in most systems its for or.

Alternators and Sub-Alternators:

Alternators alternate, whether or not the prompt is being used. What do I mean by that?
What would you guess this would do?
[[dog|cat]|[cat|dog]]
If you guessed, "render a dog", you are correct: the inner alternaters alterate like this:

[dog|cat]
[cat|dog]
[dog|cat]... etc.

[obtuse torrent]

I think it's a syntax error, possibly he was trying to create some result by alternating the neon effect with something else in each generation

[keen cradle]

how can I make the character stay far away from the camera?

[sweet rock]

You can try long shot or extreme long shot but the checkpoint just might prefer not to

[keen cradle]

it doesn't work, probably checkpoint doesn't allow that

[silk pawn]

what will happen if i put too many prompts?

is it the same for positive and negative prompts?

what is the optimal number of prompts?

[thick timber]

the images gets disfigured at the last step

any recommendations?

[prisma rover]

Good Evening SD! i need some guru help please, i am helping my daughter design a poster for a school project, she wants 2 wedding rings (pic1), i got this so far (pic2) with controlnet, i will regenerate a few more to get it perfect, my problem is, she wants the diamond to be shattered (thanos snap? :p) like pic3, and i can not for the life of me figure out how

[atomic flume]

I keep getting overly long legs

African, (solo:1.5), Masterpiece, (1990s [cyberpunk|elfpunk|sci-fi] art), beautiful lighting, front view, highly detailed painting of (alluring [wood elf|man|cyborg] techno mage with short cyberpunk hair), (wood elf ears:1.2), cyber optics, perfect face, (mature black male:1.4), beard, (wearing professional suit with tie:1.4), (detective:1.4), (button up shirt:1.2), (vest:1.2), (trench coat:1.2), combat boots, mechanical joints, robotic parts, silicone, wires, [burn fuse], athletic body, vivid details, cybernetic tattoos, (Art by Frank Frazetta and Jeff Easley), cyberpunk artwork, akira, oil on canvas, (perfect composition), vivid colors, intricately detailed, fine details,

Dystopian cyberpunk Tokyo in the background with skyscrapers, [neon lighting], neon ads, Fujifilm Provia 400X, (film grain:0.8), (full body:0.9), (walking towards viewer:1.2), <lora:epi_noiseoffset2:1>

Negative prompt: (bad-image-v2-39000, bad_prompt_version2, bad-hands-5, EasyNegative, NG_DeepNegative_V1_4T, bad-artist-anime:1.2), (nude:1.2), long legs, canvas frame, ms paint, ((disfigured)), ((bad art)), ((deformed)),((extra limbs)),((close up)),((b&w)), weird colors, blurry, (((duplicate))), ((morbid)), ((mutilated)), out of frame, extra fingers, mutated hands, ((poorly drawn hands)), ((poorly drawn face)), (((mutation))), ((ugly)), ((bad anatomy)), (((bad proportions))), cloned face, (gross proportions:1.2), (malformed limbs), ((missing arms)), ((missing legs)), (((extra arms))), (((extra legs))), (fused fingers), (too many fingers), (((long neck))), Photoshop, video game, tiling, poorly drawn feet, (cross-eye:1.3), body out of frame, bad art, 3d render, watermark, letterbox, lowres, (error body), error hair, ((error arm)), ((error hands)), ((bad hands)), error fingers, bad fingers, missing fingers, error legs, bad legs, multiple legs, error lighting, error shadow, error reflection, text, error, extra digit, fewer digits, cropped, worst quality, low quality, normal quality, jpeg artifacts, signature, username, ((error eyes)), ((bug eyes)), ((bad eyes)), bad mouth, error mouth, (error face), (nsfw:1.2), to many legs, (three legs:1.2), logo, water mark, black and white,
Steps: 60, Sampler: Euler a, CFG scale: 7, Seed: 1949123979, Face restoration: CodeFormer, Size: 520x880, Model hash: d78ca06f25, Model: electricEden_v10, Denoising strength: 0.7, ENSD: 31337, Version: v1.2.1, Hires upscale: 2, Hires steps: 60, Hires upscaler: 4x-UltraSharp



Used embeddings: bad-image-v2-39000 [b03e], bad_prompt_version2 [afea], bad-hands-5 [10ca], EasyNegative [119b], bad-artist-anime [53f1]

how do I stop that ?

[pine moth]

Hey guys and gals, im trying to workout if there's a prompt delay method that works to say, start a prompt term/word up until a certain step/% or that starts at a certains step/% of the generation, from reading the very brief guide on the sd guide the methods shown only basically start and stop prompt delays, ie it's like black and white, when one stops with the : command the next part starts... is there way to have it so a letter prompt/term that is added is just simply added in with everything else without cutting it off? Hope that makes sense, thanks

[rich crow]

if im understanding you right, you can still use that same method but just add the extra text in afterwards for the second part. i.e. "[picture of a dog:picture of a dog barking:.8]" will do just the dog, then afterwards do the dog barking at 80% of the way thru

[leaden socket]

Anyone know how to remove a split down the middle of a generation?

[rain jay]

I am creating two persons. How can I specify the characteristics of each person? (ping me)

[rich crow]

Use this regional prompt extensionhttps://github.com/hako-mikan/sd-webui-regional-prompter

[rain jay]

thanks

[rich crow]

here is a popular post that explains it https://old.reddit.com/r/StableDiffusion/comments/13360c1/allure_of_the_lake_txt2img_region_prompter/

[rain jay]

this break is a new line?

[rich crow]

hmm that thread might be outdated come to think of it

[rain jay]

i took that from the extension github

[rich crow]

the UI was really intuitive to use when i tried it last week

idk if its the same one but the one i used, you could simply type in what prompt you want for the box you specify

yeah it should be the same one

not sure if you still have to use those BREAKS (i didnt) or if the extension just intuitively appends them in for you automatically now

[sweet rock]

More info: https://stable-diffusion-art.com/regional-prompter/

They deal with two people one, image. The github example doesn't do one of those

[sweet rock]

Any good positive/negative prompts to get flatter images (as in no gradients). I've tried various vector graphic terms & "gradients" in the negative prompt but so far I haven't found a good prompt to prevent this.

[leaden socket]

svg

illistration

[sweet rock]

thank you! That definitely helped! 💖

[pine moth]

Thanks for response, I find this whole area confusing tbh lol but what you are saying there sounds like how it is described in the manual, the second part willl stop generating the first part entirely? What I'm asking about for your example would be is it possible to have the dog barking come in at 80 but as an addition to the first part of the prompt, not just like swapping, you know what i mean? like to illustrate clearly imagine it's [blue sky:yellow sky:0.5] so 50% of the way through it switches to a yellow sky but it will be night and day, the first part stops getting generated, im asking if there is a way that at 50% the yellow part is simply added to the prompt but not stopping the first part, simply an addition ?

[obtuse torrent]

@pine moth I can't get the underscore to appear in chat ...
try putting an underscore in the initial prompt ,_, and then replace it with what you want

[orchid ore]

escape it right alt+q and _

[obtuse torrent]

@pine mothfor you example ....... ,blue sky, [_:yellow sky:0.5] .... (blue sky never is stoped) .. @orchid ore if it is for me thx

[orchid ore]

yes but i think you solved it already before 🙂

[obtuse torrent]

try to lower the weight of these tags (wearing professional suit with tie:1.4), (detective:1.4) .. (if adding normal, or short legs not work)

[pine moth]

aha! I tried the _method and i think it is working indeed!

thanks

that's a clever idea

my question now is what does the double :: do ?

for example like

"blue balloons, [_:(yellow baloons:1.9)::0.5]"

or would that be more applicable if you wanted to STOP a word/term for generating, so used earlier on in the prompt?

infact, i have tried a different prompt and it doesn't seem to be working lol, on the balloons method with clearly different colours it worked as expected , at 50% yellow started to appear but on a more detailed prompt it isn't working

" detailed photo of new york city skyline ,[( _ : cinematic orange and teal, film grain, vintage, hyperdetail:1.9):0.50] " as described, i want the latter section to come in half way through but watching the preview, the latter part is being generated right from the start of the prompt, any ideas?

nvm found the error! the bracket in the second part was encapsulating the first part of the prompt but should have only started before the cinematic word, so easy to goof up here lol, thanks! hope this helps someone else too

[obtuse torrent]

for it is a syntax error ... but ...[prompt::0.5] ---> at 50% of the steps ... stop prompt...

[pine moth]

And how do you define between the :x number being the amount of steps vs percentage? Like :0.20, :0.50 0.80 will always refer to percentage amounts yeah? So if the amount of steps for the image are 10 or 140 :0.80 will be at 80%, right? If you want to to do it by step count is it as simple as it being say :20, :50 or :80?

[obtuse torrent]

For standard feature always is how a %, ...I don't know if there is a script that adds the option to specify the number of steps. (it doesn't seem difficult, it's a rule of 3 with the total number of steps, but I'm still ignorant on that subject)

[pine moth]

Ohk, percentage works fine anyway. Thanks!

[pine moth]

[(detailed skin:1.5):0.1::0.2] would this be the correct syntax if one wanted the prompt to start at 10% and finish at 20% ?

[signal cradle]

so you know how a lot of loras and text inversions all at once make the image a mess? Any way to balance that? higher CFG? or maybe lower? more steps?

[lavish surge]

How to prevent the image being divided?

[rich crow]

that's really just the result of adding all of those together. best thing you can do is tone them down with lower weightings like :.3 or something if you really want to keep them all.
some finetuning is destructive and also all finetuning intent is to change the algorithm of the model so throwing all of those loras and embeddings together makes what youre seeing sorta make sense when you keep in mind that the AI is literally just drawing numbers.

[obtuse torrent]

first pading a " ,_, " ..... then maybe it can work (never testeaded) @pine moth

[obtuse torrent]

reducing weights, ..... the solution varies because the class of LoRas and the models in which it was trained influences, since the LoRas somehow absorbs the weights of the models and when added to others LoRas ... some tags are enhanced (it intrinsically pushes some tags to the saturation). other approximation is regional prompt isolating LoRas.

[signal cradle]

does that mean 2 Loras trained on a related concept are more likely to do this (is this burning?) than loras trained on unrelated concepts?

like, theoretically if I had 2 Loras with opposite weights, they'd null?

[obtuse torrent]

@signal cradle Yes,
It is what I suppose from experience, in the same way that when you mix models with strong NSFW you begin to see it appear in each generation, and seeing loras of characters canceling each other is not uncommon either.

[latent girder]

I'm trying to make a cyberpunk mixed with dark fantasy image where there'd be creepy half-creature half-machine and hands creeping from the shadows in a cyberpunk city. I've managed to make the city background, but I can't seem to manage to add creatures or hands. I've tried using models specifically made for creatures, with inpainting and using controlnet but it takes ages for subpar results

This is the cyberpunk setting I've managed to create, which I'm more or less happy with

But I'd like to add dark fantasy creatures kind of like that but with a more "machine/cyberpunk" vibe, and I'm not sure what to do

I tried inpainting hands with images like this in controlnet but that gave terrible results, maybe I should just photoshop them in directly and then img2img to recreate the scene?

[dusky scroll]

Hy

How make a vertical images?

[pallid shale]

Hello I'm trying to make a drawing of a pumpkin in this style. I've tried a bunch of times with medieval art terms and all that but it keeps coming up with drawings that seem way too modern. Could I get some help on how I could achieve this?

[sweet rock]

I used “Medieval print illustration, flower” & got ok results. Pumpkin seems to make modern jackolanterns though

Not exactly the same for sure

[pallid shale]

Hmm yeah, it got a little better but keeps looking too modern

Thank you for the help though

[sweet rock]

A bit better

[cinder crag]

Hello, I've been using stable diffucion for a couple of weeks, but I have some problems that I'm looking to solve. Is there someone who is willing to teach me for a fee? I do not have much
note: i only speak spanish
Anyone interested please send me a DM.

[normal cipher]

Hi, I'm having trouble with poppy flowers. That term just hijacks anything in the beginning and end of prompt and I end up with a field of poppies. At the minimum, I'm looking to generate white poppy flowers bleeding/dripping red blood.

[orchid ore]

@normal cipher how about put ((poppy-flowers)) or similar, does it trigger it too?
Going to test it.

[normal cipher]

Thank you. No difference.

[orchid ore]

are those this one yes

[normal cipher]

Sometimes this:

[orchid ore]

sorry my english is bad. I dont understand what you want achieve exactly. If just single flower, or what 🙂

[normal cipher]

I want them to bleed or be covered in blood. 🙂

[orchid ore]

o.k.

[normal cipher]

I tried having a poppy field of flowers appear outside of a burning/destroyed city -- nope just a random poppy field.

So far only "vase" has truly altered the image for me. (I haven't tried much outside of what I'm trying to generate.) It (poppy flowers) seems to hijack the prompt.

[orchid ore]

oh you want vase i read it after creating this 🙂

[normal cipher]

No vase. That's way better than anything I've generated.

🙂

[orchid ore]

i used img2img contorlnet, not sure it is must. Send it here from txt2img and typed "make it bleeding"
I think it can be possible without controlnet. mmnt

[normal cipher]

Thank you. I'm going to try regional-prompter.

[orchid ore]

it needs controlnet it seems.

you can add this on edge of image and make it blurry, can look o.k. @normal cipher

[low flame]

Helloo, any idea how to get the faces better with image to image, they look horrific even with face restoration on

[rich crow]

face restoration doesnt work well with some models in my experience

[orchid ore]

Try face restoration in extra tab. With codeformer and that second gan there. But it should work with realistic faces... So dont know how it will look @low flame

[low flame]

Thanks, I searched a bit and seems like yeah it doesnt work with other models as @rich crow mentioned

[orchid ore]

expected it way worse

[orchid ore]

@low flame this isnt face restoration on txt2img page. It is on extra tab working different way. Check for codeformers and GFPgan there.

[lost crest]

for the textual inversion / embedding EasyNegative, I see people use it as easynegative but some people also do EasyNegative. What is the correct usage, or does it not matter?

[obtuse torrent]

all model has a "hash" .. for embedding you can change names, check hash too ...

[lost crest]

ty

[desert rain]

Hello, is someone free to help me RN ? I'm not english native and i struggle searchings on forums the issue i have, i can explain it with an exemple, it'll be easier to understand

Here is one of my gens (pic1) + upscaled ver (pic2) while quality supposed to be this good (pic 3), i wonder why it is so different. I check a lot of people's gens and even if i copypaste their exact settings, i get a bad quality, i use a lot of negative prompts to have accurate result but sometimes i remove them all to test and it's not better

[lost crest]

Do you use embeddings like EasyNegative?

Can share ur prompts?

annd model

[desert rain]

hm yes, i'll show you

prompts : <lyco:GoodHands-beta2:1.2>, (masterpiece), best quality, expressive eyes, perfect face, vampire, black_pantyhose++, wearing_skirt, red_hair, bats-on-background, medieval_outfit, outdoors, night sky, full moon, castle background, masterpiece, highres+, absurdres, 4K, 8K, 16K, hyper-detailed, deep eyes, stunning, sharp, 1girl, realistic,

negative prompts : (worst quality:1.4), (bad quality:1.4), (low quality:1.4), bad anatomy, liquid body, liquid tongue, disfigured, malformed, mutated, anatomical nonsense, text font ui, error, malformed hands, long neck, blurred, lowers, lowres, bad anatomy, bad proportions, bad shadow, uncoordinated body, unnatural body, fused breasts, bad breasts, huge breasts, poorly drawn breasts, extra breasts, liquid breasts, heavy breasts, missing breasts, fused ears, bad ears, poorly drawn ears, extra ears, liquid ears, heavy ears, missing ears, fused animal ears, bad animal ears, poorly drawn animal ears, extra animal ears, liquid animal ears, heavy animal ears, missing animal ears, text, ui, error, missing fingers, missing limb, fused fingers, one hand with more than 5 fingers, one hand with less than 5 fingers, one hand with more than 5 digit, one hand with less than 5 digit, extra digit, fewer digits, fused digit, missing digit, bad digit, liquid digit, colorful tongue, black tongue, cropped, watermark, username, blurry, JPEG artifacts, signature, bad hairs, poorly drawn hairs, fused hairs, big muscles, ugly, bad face, fused face, poorly drawn face, cloned face, big face, long face, bad eyes, fused eyes poorly drawn eyes, extra eyes, bad tails, bad mouth, fused mouth, poorly drawn mouth, bad tongue, tongue within mouth, too long tongue, black tongue, big mouth, cracked mouth, bad mouth, dirty face, dirty teeth, dirty pantie, fused pantie, poorly drawn pantie, fused cloth, poorly drawn cloth, bad pantie, yellow teeth, thick lips, bad cameltoe, colorful cameltoe strong girlworst quality, low quality, normal quality QR code, bar code, white pantyhose, bad pantyhose color, more than 1 body

model : breakdomain
Sampling : i try most of them but mostly DPM++2M Karras

I copypaste negatives prompts from other person so i got easynegative sometimes and it looks almost the same

[desert rain]

I'll try to screenshot and do another one

[lost crest]

but do you actually have it installed?

[desert rain]

no, is it a lora ?

[lost crest]

it is a textual inversion / embedding

https://huggingface.co/datasets/gsdf/EasyNegative

[desert rain]

I'm pretty new so i didn't know (i don't even have your upscaler for exemple)

in which folder do i need to put it in pls ?

[lost crest]

embeddings

[desert rain]

Thanks, and i need to put "easynegative" in prompt/negative prompt after that ?

[lost crest]

you put EasyNegative in your negative prompt

[desert rain]

I'll restart stablediffusion before just in case

Oh and i saw you use hires.fix while i never use it

So i'm trying with it on, to see if the issue was here

aouch

out of memory

i'll do without

[lost crest]

you also can try latent upscaler, see if that works

[desert rain]

basic Latent ?

[desert rain]

i forgot to say i have a gtx 1050 gc so i use lowvram idk if it causes bad quality, i know it take times for sure

[desert rain]

i'll try

[lost crest]

well that may explain why you cannot upscale to 1024 x 1024

this is when I disable hires. fix

[desert rain]

Almost at 50% currently, i'll show you when finished (it will be without EasyNegative and Latent)

(but next gen will be with them)

oh, finished sooner

[lost crest]

👍

[obtuse torrent]

@desert rain ...your path\stable-diffusion-webui\embeddings

[desert rain]

i got 2 files from the huggingface site

so i put them all

Ok so this time i launch with latent and easynegative on negative prompt

[obtuse torrent]

@desert rain too you need check if your model ...need a VAE ...

[desert rain]

I didn't see something like this

https://civitai.com/models/50520/breakdomain

i have these VAE

[lost crest]

The recommended VAE is "vae-ft-mse-840000-ema-pruned.ckpt".

AnythingV5 is a model

[obtuse torrent]

@desert rain

[desert rain]

oh, didn't saw the show more, mb sorry

[obtuse torrent]

you need the standar VAE ....

[desert rain]

is it automatic ? or do i need to activate something ?

after i put it in the folder

[obtuse torrent]

@lost crest @desert rain check it and how install 😄

@lost crest thanks

[lost crest]

go to settings, go to user interface and copy these

you will see this

after a restart

[desert rain]

Thanks i'm doing it rn

I think it will take at least 10 minutes to have my gen done

Clip skip 1 is good or 2 is better ?

for now

seems correct from what you told me

[obtuse torrent]

@desert rain 1 and 2 are good but results change, traditionally Anime models use 2, but not always, Model author normally recommend a clip skip

[desert rain]

OutOfMemoryError: CUDA out of memory. Tried to allocate 1024.00 MiB (GPU 0; 2.00 GiB total capacity; 1.67 GiB already allocated; 0 bytes free; 1.69 GiB reserved in total by PyTorch) If reserved memory is >> allocated memory try setting max_split_size_mb to avoid fragmentation. See documentation for Memory Management and PYTORCH_CUDA_ALLOC_CONF
Time taken: 6m 40.37sTorch active/reserved: 1713/1728 MiB, Sys VRAM: 2048/2048 MiB (100.0%)

It was at 50%

And was looking better than the finished gen i did before

[lost crest]

try upscaling from 256 x 256

[dusky sail]

I don't you can run hi-res 1024 at 2GB VRAM...

[desert rain]

Trying it rn

[lost crest]

your 1050 cannot really handle such image upscaling

I think

as expensive as it sounds, I think upgrading your graphics card would be a good idea (only if you have the money of course)

[desert rain]

I'm french and there is almost none tutorial on stable diffusion, so i lack of knowledge

[lost crest]

there are plenty of english tutorials

[desert rain]

(i'd love to upgrade)

[desert rain]

french tutorials, i forgot one word

yes it's difficult for me to understand sometimes the tutorials and know what to type in youtube

[lost crest]

https://www.youtube.com/@OlivioSarikas Olivio is very good

[obtuse torrent]

@desert rain it extension can help you TILED VAE

[desert rain]

Thank you for the channel

[desert rain]

I'll install it

thx

here ?

i'll uncheck all "hide"

[lost crest]

https://www.reddit.com/r/StableDiffusion/comments/13utcbg/color101_vae_is_released_can_provide_sharper/?utm_source=share&utm_medium=android_app&utm_name=androidcss&utm_term=1&utm_content=share_button

[obtuse torrent]

@desert rain i not sure if it work whit DDR5 and possibly you need set slicers to minimum

[lost crest]

I just came across this on reddit

[desert rain]

i have DDR3 if i'm not wrong

[desert rain]

i'll read

[dusky sail]

You can think of CLIP skip as the resolution at which you prompt is understood. Different models may recommend different settings, usually 1 or 2. It prunes the details, so if you type e.g. "red apple with a leaf on a table" and you have some skip it may get understood as just "red apple"

[obtuse torrent]

@desert rain installed it look

[desert rain]

i put this link in index ? https://github.com/pkuliyi2015/multidiffusion-upscaler-for-automatic1111

[desert rain]

Thank you, i'll remember

[desert rain]

result :

Seems a bit less noisy

[lost crest]

yea so now use something like https://civitai.com/models/16993/badhandv4-animeillustdiffusion to improve the quality of the hands

it is also an embedding

and goes into the negative prompt

[desert rain]

so i don't need the Locco/lyco better hands after that ?

[lost crest]

not really

[desert rain]

noice !

[lost crest]

textual inversions / embeddings will do the job too

[signal ice]

old hands while praying

[lost crest]

old hands?

[desert rain]

badhandv4 go in negative prompt, roger

[obtuse torrent]

@desert rain in upscallers you too can try con others too ,

[desert rain]

i don't have them for now

i don't know which file to pick

bc i tried putting the link here :

but it show me an error

and i was unable to change it

so i restarted ui

[obtuse torrent]

@desert rain you need click in Load From ...search it and click in install... change to first page and clickk in aply and restat UI

[desert rain]

Oh so it was like this screen

Simplier, restarting

So i need now to see what is my best resolution ?

[obtuse torrent]

ok ..@desert rain ..the safe (and slow way) is ..all checked ..and the slicers to the left..

not use tied diffusion...

[desert rain]

slicers ? the bar ?

[obtuse torrent]

yes

[desert rain]

like that ?

tiled diffusion is off

[obtuse torrent]

the Sizes ... low is more slow but more stable... default more fast but you need check res to no get memory errors...@desert rain ...yes diffusion is for others things and not work good in low range Graphics..

@desert rain check if you can use native 516x516 res ...

[desert rain]

Oh, okay, good to know

i can do 1024 x 512 normally

(this was my very first prompt saved in my computer)

and i know that i can upscale these resolution by 2 but after the gen is finished and send to extra

even 1024*1024 seems to works for now

55%

without upscaler i can go to 1024*1024 then and ngl except the bats it looks pretty ok

result here

good progress

[obtuse torrent]

@desert rain there exist extensions for work in others languages in A1111 ... i never search for it (native Spanish here) but they can help you prompting in you language (french?)..

[desert rain]

(yes french)

i try to learn with english, but yeah it seems good alternative if i struggle so much

this was my best gen until now (it was upscale after gen ofc)

(and i putted the Uta lora and it's always ugly even i put prompts like "red_and_white_hair"

[obtuse torrent]

@desert rain ok nice to meet you, good luck HF .. Anime channel guys can help you in details abut anime generation if it is your interest ...

https://discord.com/channels/1002292111942635562/1091193032273043516

[desert rain]

Thank you ! Nice to meet you

The help you three gave me was great

[dusky sail]

Stable diffusion doesn't really understand which properties go where. If you say "red hair" you'll probably get "a red-hair" because that's a defining property of an entity but if you say add "blue and yellow shoes" there just will be blue, yellow and shoes somewhere on the picture.

There are remedies for this like regional prompts or some img2img workflows where you sketch what goes where to guide the model a bit

[desert rain]

So should i say "dual_color_hair (white and red)" ?

Because with the Lora, the character doesnt seem to have the good hair color just with the trigger word

(One half is red other is white so a bit complicated)

[obtuse torrent]

@desert rain paste a pic of correct hair colors and the name of character ...

[desert rain]

Uta from one piece

https://civitai.com/models/5588/uta-one-piece-lora

i use this lora

i put the heigh lower

with prompts hat fit her

[obtuse torrent]

,,@desert rain,, prompt it to guide your lora ...red hair, multi colored hair, two tone hair, pale pink hair,

[desert rain]

I'll do that, thanks, sometimes it works sometimes not so i try multiple times

[obtuse torrent]

maybe (two toned hair:1.05) and... one piece anime style ..can help too (or by ...the name of original character designer, or study ) @desert rain

[desert rain]

Seems good idea

[leaden socket]

Anyone know why?

[obtuse torrent]

@leaden socket hard without all prompt ...maybe denoising strength very height ? ..,

[leaden socket]

Happens with every prompt.

[obtuse torrent]

...
what happens at each prompt??? ...what is obvious to oneself is not obvious to others.... please... and have you checked by changing the value that I indicated to you for a lower one? between 0.3 and 0.55... @leaden socket

[leaden socket]

Changing the denoising strength just decreases the randomness / influence of the models. I'm asking if anyone knows why it's happening so I can understand it.

I figured it out, for some reason upscaling to 512 creates those artifacts, so I'm guessing it's just out of the bounds of the v1

[obtuse torrent]

well 512 x 2 still low res (it depend of your VRAM of course) ...you check the effect of changing the sampling method .. @leaden socket

[silver valley]

Missing vae

[leaden socket]

I have a 4090

[obtuse torrent]

?? of my league ....

[leaden socket]

Would I get that info from the model?

[silver valley]

From the model page

[leaden socket]

https://civitai.com/models/4550/ayonimix

Used SD 1.5 VAE

So does that mean 768?

prompt at?

I'm not familar .

[silver valley]

There is the vae mentioned

[obtuse torrent]

@leaden socket here

[leaden socket]

Thanks lads, so I did download and placed it in my vae dir, still getting the same results.

[silver valley]

You need to select it

In the settings

Then hit apply, and switch models

[leaden socket]

Ah, got it

Thanks guys!

[obtuse torrent]

@leaden socket

[leaden socket]

After looking into vae, it looks like it jsut comes down to preference on the look (unless the model specifies one), would that be correct? Or is there one better than all others? Should I choose a vae outside of the default one?

[obtuse torrent]

mmm..yes.. VAE are corrections (some times specific VAE work better whit a model) @leaden socket , exist VAEs more used.. others ...clear VAE, blessed VAE, etc .. you need check effects

you can go whit vae-ft-mse-840000-ema-pruned (for realism) how a for all purpose VAE)@leaden socket

[lost crest]

what is meant by bangs, e.g blunt bangs or asymmetrical bangs?

[obtuse torrent]

@lost crest

[lost crest]

oo

[obtuse torrent]

between forehead and eye line.

[steep otter]

You can do latent coupling to use separate prompt for different regions...

[steep otter]

You can tag a few image with "blue and yellow shoes"...

[dull rune]

prompt for changing Image size

[rich crow]

That's just a generation setting, not a prompt right?

[orchid ore]

so for negative embeddings is enough to place it in embeddings and mention its name in negative prompt without pt extension? Not sure making it right, dont know how to check it.

[lost crest]

I sometimes see prompts that use _ for words instead of spaces (those words aren't textual inversions / embeddings). Does it make any difference whether I use _ or spaces?

[orchid ore]

i think it results in different pictures space or _. But should have same content. Tried it now and with space i got black woman with _ i got white woman with hands on head.
Also with _ i got more tokens taken... weird

[obtuse torrent]

.well matter to check ....... (is notation of old tagging systems) i think..is .. no difference .... _ is a neutral character ... (between tokens) ...

[orchid ore]

content should be same but picture not.

and hash of prompt which i can imagine is used as well for generating

[lethal spear]

Is there a way to do X/Y/Z prompt as separate images instead of a grid? Hoping to figure out how to get variety from one prompt, so I can crank up the batch count/size and walk away

[orchid ore]

There is very small difference between _ and - smaller than space and any of those

[obtuse torrent]

..... @lost crest ..... this interrogate clip continues tagging like this ..

@lost crest ... maybe in models based in open clip it not work more.. IDK...

[obtuse torrent]

if you need a sure way to call TI use... 1 click here ...it add the correct name to the end of your prompt..

[rich crow]

Is it _ on left and right? Like clothes
If so then they're using the dynamic prompt extension and that's their filename for a random, in this example, different clothing that they set up in a text file
If not then I'm not sure.

[lost crest]

like dynamic_lighting

[rich crow]

Might be a booru thing then, there's no spaces in booru tags

[orchid ore]

the difference is in clip i believe. I mean different results beautiful_woman vs beautiful woman. And it is clip thing i believe, not using anime.
Will try interrogate image and add _ in the middle.

[green gull]

Hi everyone, I have a question. What prompt work best to prevent letters appearing on the generated image? The attached image is an example, and I used the following prompt: megalithic tower futuristic, by Noah Bradley, gorgeous, cinematic, dynamic lighting, hyperrealistic, cyberpunk, highly detailed, intricate, majestic, vibrant, 8k resolution

Any advice I'll be very appreciated .

[obtuse torrent]

@green gull ... you can try ... text, url, web page, font ui, text, ui, watermark, username, signature, QR code, bar code,annotations, publicity, directions, specifications, data, notes, author identification, study name ...(there is a good chance that it is an effect of the seed that was used or of associating an author in the prompt).

[narrow merlin]

How to create an image?

[green gull]

Thank you! Will try those

[vague flint]

any idea how to get results to near this one?

[cinder olive]

Masterpiece, best quality, a stunningly beautiful dark fey wearing thick purple robes, fantasy, dungeons and dragons, evil fairy, detailed thick fantasy robes, detailed long skirt, detailed butterfly wings on back, detailed long black hair, detailed purple eyes, occult, detailed conservative fantasy outfit, detailed long fluffy bug ears, fully clothed, sorcerer,

Negs - ((Big Breasts, NSFW, lewd, risque, nude, lewd, sexy, sex, sexualised, sexy outfit, sexualised outfit, see-through clothes, boob window, revealing, revealing outfit:1.2,)) hat, laser, lightsaber, lowres, bad anatomy, bad hands, bad faces, text, error, missing fingers, extra digit, fewer digits, cropped, worst quality, low quality, normal quality, jpeg artifacts, signature, watermark, username, blind, bad eyes, ugly eyes, dead eyes, blur, helmet, wings on head,

How do I get it to properly clothe her? X.X I feel like I've used every prompt possible to get her to cover up.

[distant dune]

Be more descriptive with your prompts. Rather than just saying "detailed this", try to describe a particular feature of that item. Here's a list of conservative style dresses:
A-Line Dresses Shirt Dresses Wrap Dresses Maxi Dresses Tea-Length Dresses Turtleneck Dresses Empire Waist Dresses Shift Dresses Kaftan Dresses Peasant Dresses Sweater Dresses Midi Dresses Long Sleeve Dresses

You can also try including "cowl" or "square neckline" in the description of your dress. Things like that.

Also, bulky negative prompts can sometimes be counterintuitive. I would suggest reducing parts of it and/or trying some negative embeddings.

You could also try inpainting the parts of the dress you want to cover as well.

[brittle geyser]

Any suggestions how to block SD from generating these weird forehead.. jewelry.. helmet? bits

[cinder olive]

much better, thank you

[obtuse torrent]

a way to solve.... use Img2Img ... Inpaint ...mask bare skin ... Just resize (latent upscale) ..Only masked ....Denoising strength 0.5 - 0.7 ... change prompt to .. only dress descrition .. IE "Masterpiece, best quality, wearing purple dress, (covered bust, collar bone covert, long sleeves dress:1.1), detailed long skirt, fully clothed,"

[cinder olive]

👌

[obtuse torrent]

@cinder olive ... the selected model also influences when doing the inpaint..., ideally use the original, or the Inpaint version of the original ...

@cinder olive

[cinder olive]

Thank you for the advice.

[steep aurora]

Does anyone know of any way to do a random pose in controlnet?

[distant dune]

HAve you tried the OpenPose editor?

[steep aurora]

I have it installed. I have a bunch of poses I've downloaded in a folder, and I was hoping to choose randomly from that folder. Does openpose editor do something like that?

wanting to do batch generation of a dozen images at a time

each with a different pose.

Controlnet has the batch option, but it isn't random and it doesn't do anything with dynamic prompts, which I was hoping to do

[sweet rock]

3d openpose extension has a "random pose" functionality. Its as advertized though, random.

[steep aurora]

Not really what I'm looking for, but thanks!

[distant dune]

I don't know how to make it randomize the selections from the cn-batch folder. It always inputs each pose in the same order per batch. This is all I could find online thus far, but it's an empty post:

https://www.reddit.com/r/StableDiffusion/comments/11z47v7/comment/jdcl0d5/?context=3

Fixed link

[steep aurora]

One of the comments says this: "the regular batch process tab for img2img will do this. The trick is to use the poses as the input images and check the box to "skip img2img processing..."" I'm going to test it now

hmmm, looks like it doesn't work. Must have been patched in controlnet 1.1

since they added batch processing into 1.1

Aha, I think I figured it out. I found a way to do *close enough* to what I was hoping for. It's not random, it goes through each pose in order, but it works with dynamic prompts, which is what I really wanted.

[steep aurora]

In the controlnet settings, there's a controlnet setting that says "Increment seed after each controlnet batch iteration"

if that's turned on, it iterates seed after each generation, making it work with dynamic prompts

[distant dune]

Ah, cool. That's good to know!

[outer plaza]

Is it normal that, without specifying something about the legs (shorts, pants, shoes, ...), square images are more likely to generate characters without legs? It's like I'm using some Facebook AI

[rich crow]

yes, just a guess but it sorta makes sense if the LAION model that SD 1.5 was trained on was all automatically center-cropped. noone has time to manually crop millions of images to 512x512 lol. this would lead to the head and legs being cropped out of some generations of people

[orchid saffron]

How do I write prompt So that I can ask to remove text after cetain steps

[kind tulip]

are there reliable ways to prevent this kind of adult-child thing? i want them to look like they're all actually real humans lmao

not tiny adults

like he looks real

just removing robin williams from the prompt does wonders, but, man

[distant dune]

Could try img2img/inpainting the faces. Could also try regional prompting which would give your more control over each section or individual.

[kind tulip]

just might need to add better fine-tuning data next time i'm training the model

it swung the other way 🤣

[obtuse torrent]

[from:to:when] is the general syntax ... this the feature https://github.com/AUTOMATIC1111/stable-diffusion-webui/wiki/Features#prompt-editing ....... you can use ... from = "a prompt" (transitory) ..... to = (nothing) ... when = 0.xx (xx is % of total of steps) ....IE if total steps is 40 ..... then ...[red ball::0.60] count how "red ball" the first 24 steps (40*0.6)... and count how (nothing) the last 16 steps.

[mint rover]

I have a bunch of pixel art icons that I've already upscaled using nearest neighbour, I'm looking to do an img2img that just kinda fleshes it out more.
What kind of stuff should I be using here? @ me

[tired vigil]

DEAM

[orchid saffron]

Awesome thanks 🙏🏼

[brazen roost]

Hi guys, how can I do this image in sd?

IIRC, there's a specific style to these kinds of images

where the pixels kind of just glitch out down

[rich crow]

Might have to make a glitch art lora. Or try glitch art in prompt or some specific glitch art keyword+controlnet

I wonder if you could also just generate a picture as a base composition and put it thru whatever program that glitch art artists use to produce it to begin with

[pearl belfry]

Omg, getting consistency with long hair is impossible

Also it seems like a lot of anime models are heavily biased towards 'flaring' long hair out, as if it's a photoshoot..

[obtuse torrent]

@pearl belfry
Any examples to share (desired image, used prompt, bad result obtained, etc) or is this just a general comment?

[pearl belfry]

Ugh... I'm using the Meinamix model right now (it has a really nice style), though some of these bias seem quite widespread. Three biases I've noticed that are very difficult to curtail: (1) long hair seems to 'flare out' as if blown by the wind or something. Gives a dramatic, artificial look, not naturalistic. (2) Sidelocks. I can't get rid of them! (3) There's almost always hair over the front of the shoulder. Can't make it stay in back.

I want something neater and more practical, whereas this kind of hairstyle looks more 'photoshooty' if you know what I mean. Also tried single braid and french braid, and you get similar issues

Seems like long hair in general is just very difficult to control...

[obtuse torrent]

plis prompt in text format :D? @pearl belfry

[pearl belfry]

Negative: (grey hair:1.3), (sidelocks:1.3), hair in front, collar, easynegative,  bad-artist-anime```

[obtuse torrent]

Possibly it is influenced by the model, I will check in others.. ok

[pearl belfry]

counterfeit and divineelegance don't fare better. Have already tried them (also they just look worse in general)

[obtuse torrent]

it still bad? ..is same prompt in other model ..@pearl belfry

[pearl belfry]

So you can get particular images that fix one of the issues, but it seems like a small proportion of the overall results

See if you can get the hair to be brush back over the shoulders instead of partially in front

I can't find any words to communicate to the AI to do that XD

any hairstyle

[obtuse torrent]

..... (ok but nudes are banned for the bot here..against rules @pearl belfry

[pearl belfry]

trying to control sidelocks seems basically impossible

It shouldn't show anything bad if you did 'extreme close up'

you can do clothes, it doesn't matter

I was doing naked since you apparently want to do that for training

Like I'm basically just trying to make a character that looks naturalistic and not like a freaking runway model

[obtuse torrent]

@pearl belfry yes is really hard ..cheeking hair styles.. no consistent results .. I'll let you know if any tag appears that is consistent

[pearl belfry]

Considering how important hair is (maybe the most important), it's very annoying how difficult it is to control

Feels like I'm forced to do short hair, since that incidentally resolves all three issues

I assume these are just anime issues and not with realistic images

[obtuse torrent]

@pearl belfry ... (((fully viewable shoulders and collarbone))), ((sliked back hair)), ((short sidelocks)), (((bare shoulders, collarbone, no hair falling in front)))... a little more consistent but
adds a tendency to shots "from above"

@pearl belfry long hair, has a tendency to view from the side or from above (to show it off), or to let it fall in front of the character or spread it out like it's blown by the wind...also adds weight to the sidelocks making them longer. ...😮‍💨

[sweet rock]

Any of you folks using the "tiling" option much? I've got some questions about it. In short, whenever I try to use it the results are way too simple & cartoony...

[sweet rock]

also any good prompts for getting images like https://i.huffpost.com/gen/834435/images/o-CANDY-CORN-BUZZ-facebook.jpg (not necessarily candy corn, just piles of stuff)

[shut salmon]

Are there any docs on parameters for SD ?

[desert walrus]

Let me pick some brains.

How would you go about creating a rimlight around her head and hair?

I tried a couple different methods with inpaint, but I am not happy with the results.

[rich crow]

inpainting wouldnt work to begin with, no matter how many times you add noise to the image and diffuse it, a purple background isnt going to turn into bright white light

edit it in photoshop, can easily just mask her and add a feather that has white color, inpaint should work after that.
or just add the rimlight in photoshop lol

[sweet rock]

Maybe this'll work (untested) https://www.youtube.com/watch?v=_xHC3bT5GBU

[distant dune]

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

[quiet zodiac]

@desert walrus What if you add it to the prompt? "lit from behind"

@pearl belfry Try an image guide with a picture of a person wearing their hair how you'd like.

[quiet zodiac]

@sweet rock This kind of gets close...

a top view extreme closeup, scattered across a desk, thousands of small glowing mechanical science fiction parts, piles on piles, overlapping parts

[desert walrus]

I need too close of a result for that, maybe in early stages + tile controlnet

hmmm

For now I managed to get it done with clipdrop light tool

[sweet rock]

Thanks Atom! I figured out a bit but you have some great prompt words in there I hadn't thought about!

[mental geyser]

A breathtaking celestial landscape with a nebula-filled sky, a massive supernova exploding in the distance, a solitary astronaut floating weightlessly in a spacesuit, an astronaut's visor reflecting the awe-inspiring cosmic spectacle, evoking a sense of wonder and insignificance, Illustration, digital art, --ar 16:9 --v 5

[quiet zodiac]

Here's my paraphrase of that...

((floating weightlessly)) in the dark vastness of space a solitary astronaut's visor reflects an awe-inspiring cosmic spectacle of a massive breathtaking supernova exploding in the distance evoking a sense of wonder and insignificance, illustration, digital art, crisp vector logos

Floating weightlessly doesn't seem to take and I wonder if it conflicts with the request for it to be an illustration? Or perhaps I need to find the right seed?

[tardy cedar]

Hey, would anyone have any insights to the prompt used to make something like this? https://www.tiktok.com/@pxl.pshr/video/7237729089002687790?q=pxlpshr&t=1685762664308

[quiet zodiac]

That looks like video converted to an image sequence with each frame used as an image guide. I would guess no animated seed, and a prompt something like...
monster mash vampire rave with a red glowing eyes

[runic osprey]

any suggestions to improve this prompt? (I'm pretty happy with the results, but wouldn't mind slightly better or more varied results). Mainly just wondering if there's a list of good things to add to any prompt/negative prompt to improve them

a __colors__ cute __types__ type __animals__ pokemon, fantasy creature, 3d animation, masterpiece, high quality best quality
neg: ugly, blurry, hand drawn, complicated, detailed

[quiet zodiac]

It certainly works. What's the significance of the double underscores in your prompt?

[runic osprey]

they're wildcards from dynamic prompts, just chooses a random word form one of my lists

[quiet zodiac]

So, that's a function from some webui? I'm running SD locally.

[obtuse torrent]

feature de a1111

@runic osprey ...maybe color-eyes for color eyes, actitude/emotion, (sad, happy. angry, etc) ... pose (maybe not work)

[runic osprey]

it's an extension for automatic1111

[cinder olive]

I'm trying to get a young boy with hair like this or shaven,

But no matter the prompt it always comes out like this;

Masterpiece, Best Quality, Taru Bisura, a plain but kind boy, detailed black eyes, short shaved black hair, old fashioned outfit, twelve year old, ninja, perfect, solo, solo focus, Naruto Style

Negs > Big Breasts, NSFW, lewd, risque, nude, hat, laser, lightsaber, lowres, bad anatomy, bad hands, bad faces, bad feet, text, error, missing fingers, extra digit, fewer digits, cropped, worst quality, low quality, normal quality, jpeg artifacts, signature, watermark, username, blind, bad eyes, ugly eyes, dead eyes, blur, helmet, eating, drinking,

Any advice?

[quiet zodiac]

If you're trying to match that image, supply it as a guide and simplify your prompt.

anime boy with brown hair

Try dropping "Naruto Style" that kept giving me spiky hair.

If you like the basic body from your second image, even a crude composite can work as an image guide to help preserve some of the look.

Try specifying the name of the actual hair style.

anime boy with brown pompadour hair

[outer plaza]

Why it's so hard to have a driver in a car? Context: previously, I was trying to add a driver to a realistic photo and failed miserably. Now I'm trying to do the same but with anime style, using reference only ControlNet to get specific characters. Reference is: <#🍥｜anime message>
I'm being able get good enough passengers, but no matter what, I can't get any of them driving. Is there some magic prompt that will make it sure any of them will be driving the car? Someone has the magic word?

[sweet rock]

Drive a car, play an instrument, hard things for ai to portray

[outer plaza]

...using a soldering station, working as a telephonist. I'm always choosing to portray things it has trouble creating
Trying with ControlNet + Regional prompt. ControlNet alone already ballooned the time to create the image to 14 minutes, hope it takes less than 1 hour for this attempt

[outer plaza]

Progress. And it was FASTER, somehow (just 9 minutes)

[silver valley]

like this?

[outer plaza]

I'm not even that specific, but I would accept it

[silver valley]

Prompt was:
1girl, black hair, in car:1.2, steering wheel, tanktop, skirt, tight, amateur board, side view, steering,

[outer plaza]

I'll try it when this batch ends, so in 20/30 minutes I should have something (I need something better than 1650 Max-Q)

[silver valley]

ah okay what do you make?

batch upscaling?

[outer plaza]

No, one batch with 3 images take this long

[silver valley]

whaaat

what are your settings 😮

and gpu

[outer plaza]

For this one, it's 24 steps of DPM++ SDE, 768x768, ControlNet reference-only and Regional Prompter. ControlNet decimates performance. GPU is a Nvidia GTX 1650 Max-Q 4 GB (now 50 W instead of 35 🔥 )

[silver valley]

wow and it works?

hope your using --xformers --medvram

[outer plaza]

It works 👏 (I've disabled ControlNet and Regional Prompter to make it faster)

[outer plaza]

--precision full --no-half --lowvram --opt-split-attention --always-batch-cond-uncond --xformers

[silver valley]

i would try without --lowvram instead use --medvram and remove --alway-batch-cond-uncond and --opt-split-attention

should be faster

[faint sluice]

So, guys, I don't have much experience working with SD.
Tried to create an image with adventurers standing on a mountain looking at foggy dungeon entrance with red eye in the dark of entrance, but that's the closest I got to what I asked.
Any advices? Or idea for the prompt is too complex?

[outer plaza]

Many things won't even start without lowvram, like ControlNet would be unusable (at least it was the last time I tried)

[silver valley]

ah okay, yea controlnet is vram hungry :/

[silver valley]

that looks very good

what do you have in mind that could be changed?

[faint sluice]

I want eye to be inside a door. Like a monster peeking out

[silver valley]

ah okay

[faint sluice]

Basically this, but a bit different

Maybe asking for foggy entrance + monster in the fog will work better than asking for red eye in the entrance, hmm

Doubt it'll work though

[outer plaza]

The image has its problems, but it's something like this the goal. Now I've got this one, I think I can inpaint it until it's good

[faint sluice]

@silver valleythese seem relatively close. I guess I'll move to 'img to img'

[silver valley]

Looks good, you could now inpaint the red eye

[faint sluice]

I decided that I'll do silhouette instead of eye

[shell pilot]

Which models and styles can be selected for prompt words？

[desert walrus]

What prompts do you guys/girls use if you want to keep the whole image in focus?

I tried sharp photo, everything in focus, small aperture and negatives like depth of field, blurry, bokeh.

Any tips?

[orchid ore]

have you used word photo in your positive prompt? @desert walrus ?

[desert walrus]

Does that add depth of field? I'll kick myself

it's so logical 😄

[orchid ore]

i think you can expect it will try to emulate photos.
My eyes are bad, just try without photo word. Or try if for example word oil painting make everything sharp.
Just tried "House in ski resort, in summer" and it seems all sharp, but having not best eyes.
I have other issue make it look like tilt shift....

[faint sluice]

@silver valley I think I got it

[teal escarp]

what can i put in prompt to make it not being a card

[desert walrus]

That might be the model you are using

[teal escarp]

no

its for real life images

[faint sluice]

Maybe you have some style selected?

[silver valley]

Nice one!

[desert walrus]

What I mean is the checkpoint you are using. With that prompt you should not get a picture of a card with a normal checkpoint.

[teal escarp]

i dont have any

[orchid ore]

how about extensions?

[teal escarp]

i dont have any

[obtuse torrent]

@teal escarp Did you try many times? maybe it's just a matter of the seed or Lyco used,,,, negative: CG, card game, copyright, legend, url, web page, text, signature, identification, (username),

[ripe rain]

guys if you zould try to do elon musk sitting on mars in an armchair, looking straight into camera ... what would you use as prompt?

[pulsar magnet]

Hello, what are the best prompts for a wide angle shot, so i get a little bit more of the background?

[pulsar magnet]

Try 2 picture: one with elon musk sitting in a chair and the other one a picture of mars surface

and then with controlnet depth or canny model

or maybe depth leres and remove background

or remove background from elon with photopea,gimp,photoshop

[sweet rock]

Depends on your model how easy it is to get that. Wide angle, long shot etc. if you’re still having trouble maybe use in paint sketch or regional prompting (or use a model that isn’t so portrait-skewed 😉)

[pulsar magnet]

hmm good idea

ty

[honest perch]

how do i fix the face? prompt: A young woman looking to the camera wearing a pink polka dot dress, depth of field, photorrealistic, 8k uhd, realistic, hyper-realistic, I Ketut Soki

[ripe rain]

can somedbody suggest a good video where it is explained how to make chatgpt generate prompts or something like lexica but with less naked women (if possible ) 😂

[quiet zodiac]

@honest perch Try declaring the shot first. Perhaps...

facing forward shoulder shot, A young woman wearing a pink polka dot dress,  depth of field, photo realistic, 8k uhd, realistic, hyper-realistic, I Ketut Soki

[honest perch]

the problems are the eyes

[quiet zodiac]

Here are some negative prompts I use for more realistic people.

tattoo,moustache,cross-eyed, deformed iris, deformed pupils, 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, bad composite, poorly masked, nude, naked, text, words, wristwatch, earrings, jewelry, makeup, hat, helmet, horns ,headphones, earbuds, bows, medals,logos,toddler, youngster,children, kid, boy, girl, cropped, out of frame, worst quality, low quality, jpeg artifacts, ugly

[honest perch]

the left eye looks weird

[quiet zodiac]

What model are you using? Try using another...
It's not uncommon to photoshop a final. You may have conflicts between your positive and negative prompts.

[honest perch]

rpg4.0

[quiet zodiac]

For people, I have good luck using RealisticVision 1.4. It also works great for tech devices and interiors.

[honest perch]

how can i make a long image?

[quiet zodiac]

Change the aspect ratio. I use 576x768 in those cases, and often throw "head to toe" into the positive prompt if I want the full figure in frame.

[crimson kettle]

Hey guys, probably gets asked a lot here but English aint my first language so the web ui section was confusing. If I want to permanently exclude a negative prompt... do I do {} or [] or the () whatever it was?

[honest perch]

any other type of cameras for get better result?

[obtuse torrent]

permanently exclude a negative prompt..??? ¨{} and () are for positive emphasis ...... [] is/was for negative emphasis (in UI whit single prompt box) .. ,,,,and now used too in some features

[crimson kettle]

So... if I was to want to perma excluded "Red clothing" how would I write it, is it enough to do [red clothing] and the image wont generate it? or? Sorry for the confusion thank you for helping out

[obtuse torrent]

@crimson kettle in A1111 you have 2 text boxes ... the first one you put what you want (positive prompt)... in the second what you don't want (negative prompt) .... but this is normally not permanent it only works as long as you generate with those prompts

[crimson kettle]

Right, so what is the point of putting in negative prompt lets say... (worst quality, low quality:1.4), I found it from someones prompt, how does it work then, do some generations wont have worst quality/low quality, or literally all will have?

[obtuse torrent]

@crimson kettle it work because is a help o guiance, in some point, when the original image sets were processed, they were tagged (indexed, references were made) those same "tag" are used to look for reference images both to add to the generation and to exclude... in the negative you ask to exclude classified things like "low resolution" for example

[crimson kettle]

I think I am just too stupid for this, I dont understand haha...

[obtuse torrent]

for most people they don't consider "why it worked" or "how it works" important... they just need to know "how to use it"... it is used by putting what you want in the positive prompt, and what you don't want in the negative

[crimson kettle]

So just to clarify as example. Positive: (red car), car, vehicle. Negative: [green car], truck, airplane

What does this do? Will a green car ever be generated or?

[obtuse torrent]

@crimson kettle normally not because you are asking for a red car.... (and other reasons of mathematical importance... such as no other colors are mentioned, it is not a very long prompt, etc)..... but there is still a slight possibility that someone has misclassified a green car as red... or that the model takes as a guide "a car" and not "a red car"

[crimson kettle]

Oh right, okay okay, so we basically just hoping and adding more hope in hopes of getting a better result

[obtuse torrent]

@crimson kettle in your example ..vehicle...is a call for misssmatchs .. ... still (motocycles, trains, bykes, etc)

[crimson kettle]

Understood, aaah right, makes sense actually thanks so much my dude!

[obtuse torrent]

@crimson kettle There are abstract concepts that influence, such as "model weights", "order of appearance in the prompt (relative weight), "number of related tags (reinforcements, contradictions, ambiguities)", "tendency to cross attention", "structure of the prompt", "references (authors, styles, studies, communities, etc)

[crimson kettle]

Basically its going to take a lot of generating heh... but thanks so much you cleared it up a lot

[obtuse torrent]

@crimson kettle start whit basic ...
ask yourself "what do I want", and then ask the community "hey guys, how do I get this" ..and then you use that knowledge and the classic "trial and error" 😄

[honest perch]

any good models for analog images?

[obtuse torrent]

@honest perch Do you mean, realistic models with which to generate images similar to those taken by analog cameras?

[honest perch]

Yes of course

Any good one over there

[obtuse torrent]

@honest perch IDK much ..maybe .... deliverate, realistic Vision ... you can ask in https://discord.com/channels/1002292111942635562/1072238304042438758 for a orientation ...ask too for cameras models, and effects .. to add at prompts

[tired vigil]

Which website do you use?

[honest perch]

how do i fix it?

[silver valley]

There is analog diffusion or try realistic vision

[tulip ibex]

hey guys looking for a prompt to help with vibrant colors as sd is washing the colors out really quite bad and i dont know what prompt to use to fix this i tried contrast and high_contrast but i dont know what else to suggest it atm

[silver valley]

Thats mostly if you dont use a custom vae file for the model

Here you can compare it:

[rare ember]

I havent had any luck changing the theme of an input image, is there any example of how to do that? As in, I put an environment img and have SD change it to make it look desertified, snow theme/xmas and so on

[orchid ore]

probably you mean controlnet @rare ember

[delicate comet]

I am trying to make some interesting looking animal hybrids. Does anyone have any suggestions? just using something like cat peacock hybrid doesnt work; using something like manticore or chimera doesnt seem to work either

[orchid ore]

try cat crossed peacock for example, + some sauce like photo-realistic and so

@delicate comet and you need some patience to get desireable result.

[delicate comet]

@orchid ore Ill give it a shot, thanks. I am fairly new to learning what works with this AI

[tulip ibex]

honestly i dont know much about vaes i need to research on that

[quiet zodiac]

@orchid ore What model did you use to produce that image? I've been trying to make a one-horned animal that is not a unicorn and it seems to stump SD. I'm going to try your "crossed" suggestion, though.

[orchid ore]

i just closed it. Put image in the png info in A111. I will come in A1111 much later.

[quiet zodiac]

I'm not sure what that means, I've never used A1111. I checked metatag file info on your image. It appears to be blank.

[orchid ore]

o.k. it was coincidentaly on realism engine, but should work with any model @quiet zodiac

[obtuse torrent]

@quiet zodiac parameters used by @orchid ore

photo-realistic
manticore crossed chimera
full body view
Steps: 30, Sampler: Euler a, CFG scale: 7, Seed: 84379541, Size: 768x512, Model hash: b513c6287d, Model: realismEngine_v10, CFG Rescale φ: 0, Version: v1.3.1

[quiet zodiac]

Cool, thank you for fetching that. I'm using MLOPs inside Houdini to run Stable Difussion.

[obtuse torrent]

@quiet zodiac what was the exact result you were looking for? (curiosity in 95%)

[quiet zodiac]

I'm trying to make a unique fantasy creature mentioned in a sci-fi book. It's a shaggy beast with one horn, but not a unicorn. I've been using the openjourney model and just wondered if there was something better for fantasy creatures.

The fur is supposed to be slithering snakes, but I can't get that either...
Also extra limbs seems to be a problem no matter what negatives I use.

I'm pulling down the realismEngine, but I've never converted a safetensor to Houdini's file format. There is a tool. fingers crossed.

[young cipher]

(simple white background:1.4),(white_theme:1.3),straight face,
(Turn on the light:1.4),
small breasts,
1girl,solo,(nude:1.3),breasts,collarbone,shoulder(straight hair:1.5),stand,

My picture has too many shadows on my face.
why

[obtuse torrent]

@quiet zodiac oh ..then it not work too

[quiet zodiac]

Cool. I have been looking over some references to the mythological Siberian unicorn, which is more dinosaur-like. In the book, the creature is also a demon.

@young cipher Try adding.. "well lit". I think it's wierd that people don't try actual sentences and just offer a list.

[obtuse torrent]

@quiet zodiac

@young cipher ...maybe "Bright Front face Lighting" can help a lit ..

@quiet zodiac lol

it throw a saurian ehit long hair ... 😮‍💨

[quiet zodiac]

Here's the prompt I've been working on for the creature...

under a turbulent purple sky, ((a shaggy quadruped demon from hell)) with (a single ornate horn) stands alone in a tall forest, tentacle like fur, fluffy tail, red glowing eyes

And an extensive negative list...

collar,saddle,dog,horse,extra limbs, gross proportions, malformed limbs, missing arms, missing legs, extra arms, extra legs, extra foot, fused fingers, too many fingers,extra horns, sketch,jpeg artifacts,duplicate,amateur, poorly drawn, ugly, flat, mangled, bad, disfigured, low detail, cheap, cropped, cut off, part missing, clipped, text, words, logo, font

[obtuse torrent]

@quiet zodiac oh i see ..
I see that I can't help you too much because all my models are anime mixers, and they are based on tag systems... I'll see what it throws at me

[young cipher]

"Bright Front face Lighting"
OK i see
I know
thank you very much

[obtuse torrent]

@quiet zodiac

[obtuse torrent]

@quiet zodiac
You can also search civitai for specific models that handle mythical creatures, but I don't know if you can do them remotely, but maybe reading the prompts can help you. use it to view proms online https://www.metadata2go.com/view-metadata

[quiet zodiac]

Thanks. I'm using MLOPs 1.0 for now. 2.0 is in the works and supports the Lora models, but I haven't taken the dive to update yet. I'm pulling a few models/CKPT down now. Maybe I'll have better luck with them.

[obtuse torrent]

@quiet zodiac i can copy here the prompt if you need someone ..

[quiet zodiac]

Are you using the "crossed" technique Bernix mentioned? That's probably what I'll try next once these models come down and I can convert them. Not on fiber, so it's taking a while.

[obtuse torrent]

@quiet zodiac 8k CGI, promotional Digital art, digital ilustration, of ((a shaggy quadruped predatory dinosaur shaped demon)), (a single ornate horn in forehead:1.2) stands alone in a (tall winter forest:1.05), (fur made of tentacles:1.1), (fluffy tail:1.05), red glowing eyes, full body, under a turbulent (purple sky:1.05),...........................................Steps: 36, Sampler: DPM++ 2M Karras, CFG scale: 6, Seed: 595498852, Size: 700x700, Model hash: 0a03c339c7, Model: divineelegancemix_V5, Clip skip: 2, ......... Used embeddings: bad-hands-5 [10ca], bad_prompt_version2 [afea]

[crimson patio]

What is the best way to make two different characters in one photo interacting with each other? Often proms do, for example, two characters with blonde hair, and there was only supposed to be one such character.

[obtuse torrent]

@crimson patio you can try extensions ...regional-prompter ....controlnet ...Composable LoRA ....Latent Couple .... cutoff ...

[crimson patio]

And which of these is the best and easiest for a novice?

[obtuse torrent]

@crimson patio cross-attention is an intrinsic part of the AI ​​learning model, so a word not only has its meaning, it is also influenced by the rest of the prompt, and is also influenced by the contexts in which that word was related in the original dataset , a strong concept leak, to the rest of the prompt, such as the colors,......... the only ways to deal with it are retrains (models, LoRa, Lyco, Ti, etc) and these tools (which are obviously not easy for a beginner). .........At the lowest level of knowledge these are just tips: keep the prompt short, be clear and concise with the description, use emphasis "()" to highlight things the AI ​​seems to "forget", add negative prompting to try to exclude what it doesn't. is desired.

[crimson patio]

Ah, ok!
I have noticed that many people when they try to describe two different people in a picture often write like this: (1girl, long hair, blue dress, purple eyes), (1man, tall, black suit, red eyes) - Does it make sense to write this way? Do these parentheses separate it somehow for AI? I'm asking because I've seen that some people write it this way.

[obtuse torrent]

@crimson patio
that helps but not in all cases, unfortunately, in that case it makes it easier for the AI ​​to recognize that there is a man and a woman, but it will be almost as difficult for it to be who is who if both are of the same sex ....naturally the AI ​​gives more importance to what it "reads" first at the prompt the "()" works as an emphasis each pair of "()" means "what is said inside is 110% more important than it normally would be "

[green panther]

Hi everyone! Good day. I need prompt help. I want generate images that have a particular height and width, how to give specific width and height into a prompt?

[obtuse torrent]

@green panther in A1111 ? web User Interface ?

[green panther]

Web User Interface. I'm talking about here in discord bots @obtuse torrent

[obtuse torrent]

@green panther
how to specify it ,,,,,,it changes depending on which interface you use in local mode,,,, and it's different with remote requests too... for discord bots ..IDK ..each one
has its own format...
I have never used a bot on this discord, I don't know if they are connected again..

@green panther reask in anime chanel ..those guys know everything https://discord.com/channels/1002292111942635562/1091193032273043516

[sudden oracle]

Has anyone been able to consistently have cat tails actually attached to their character? Most of the time mine are either showing off screen slightly like they are attached elsewhere, or they are on the ground next to my character lol

[quiet zodiac]

@crimson patio I've had some luck specify ethnicity in the case of two people. Also don't just use man and woman, make them something by giving them a profession or occupation.

(a 36 yo. syrian husband with black hair in a suit) holds hands with (a 32 y.o. irish mother with red hair wearing a silver sequin dress)

This is a way to add age to them as well, making them husband and wife. If you use the term "mother" make sure to add children or child to your negatives to prevent the kids from intruding on your scene.

@sudden oracle Duplicates and mis-matches might be related to exceeding the size limit of the model your using. Try reverting back to 512x512, or keep your width/height in the powers of two.

[sudden oracle]

Would upscaling cause this to happen more often as well?

I run at 512x512, but I upscale x3

[quiet zodiac]

Hmm..I don't think so, by the time the upscale is issued, the generation is complete.

[sudden oracle]

Thats why I am confused haha

I'll figure it out 🤔

[obtuse torrent]

@sudden oracle (in anime models you write "cat" and you get a catgirl, sometimes you write tail and... cat girl ) 🤔 ...

[sudden oracle]

I am writing cat girl and cat tail

That might be why then

[compact meadow]

Anyone have any idea how to get different eyebrows or even different eyes in any of the digital art style models?

Everyone got them same eyebrow type

and the same almond eye shape

it's always these same shapes for eyes and eyebrows just in different styles

Like see this:

same face everywhere if it's a digital art based style

ping me if you've had any success generating anything else

because this tapered eyebrow isn't the only eyebrow that exists unlike what all these civitai models think

[obtuse torrent]

Do you know a good workflow, to add a background to my character (who has the background removed) in A1111..

[quiet zodiac]

@sudden oracle You could try something more sentence like...

A cat girl whose tail wraps around her feet

or

A cat girl with a fluffy tail

Try with instead of and?

[sudden oracle]

Oooo thank you!

I was trying to keep it simple, but I will try that!

The second one helped A LOT

[quiet zodiac]

@compact meadow Try using an image guide with a face containing the features you're seeking.

[compact meadow]

What image guide

[quiet zodiac]

SD supports image guides and masking. You supply a reference image to "guide" the final result. I'm not sure how the web interfaces handle that. It allows you to blend in features from a photo.

[sudden oracle]

Some still have either floating tails or extra tails, but a good bit of them actually have good attached tails

[ripe rain]

guys if you would try to do a realistic picture of konfuzius, how would you prompt it? Please

[quiet zodiac]

I know I sound like a broken record, but use an image guide. Also try the other spelling variation for his name to see which one resonates best.

facing forward, a well lit realistic shoulder shot of chinese philosopher confucius in an inviting park setting

[obtuse torrent]

"raw photo of ..... " is the more basic ...@ripe rain (if SD AI know about konfuzius or any other concept) .... or go to.... Img2Img

[ripe rain]

thats all you typed?

[ripe rain]

so youre sayiung if SDknows about confucius then i can type raw photo of... ?

[trim sigil]

I am trying to get females with their hair covering a portion of their face similar to this screenshot - but I am struggling with almost any model I use. I am working with Fantastic Mix 2.5D right now but pretty much every model I've tried gives me bangs at best. 🤔 Even if I put bangs in negatives. I'd also like tight curly hair and usually again - get waves at best.

The prompt I am using right now is:

(masterpiece), (best quality:1.2), absurdres, [intricate details:0.2], solo, ((beautiful detailed eyes)), (detailed light), depth of field, 1girl, ((crimson red hair)), ((long thick curly hair)), hair covering right side of face, (short pointed ears), ((green eyes)), detailed eyes, freckles, blush, large breasts, cleavage, smile, bamboo forest, japanese shrine, outdoors, forest landscape, chinese fantasy clothing, intense shadows, chubby, curvy, lora:add_detail:1.3, mature, sexy, saturated colors, dramatic lighting, (pale skin:1.5), green eyeshadow, juicy lips, portrait, lora:LAS:1.25

[obtuse torrent]

raw photo is for a standard photo ...all photo are realistics...@ripe rain .....(assuming you are looking for a "real" photo... you can also declare "a paint".... "digital art"... "a CGI"... etc. (in those other cases adding terms like realistic, photorealistic, hyperrealistic, makes sense, because you are declaring a style)... if SD IA know about a concept it work ... (some ..celebrities, historical figures, locations, etc)

@trim sigil try "hair over one eye"..

[trim sigil]

Thank you, will do!

( That is already helping, thank you so much 🙇‍♂️ )

[slate axle]

Can someone explain why I am getting face like this no matter what?
using controlnet with more focus on prompt.
prompt: ||a demonic woman with horns and wings on her back, standing in a dark red fog, (art by Anne Stokes:0.5), dark fantasy art, concept art, gothic art, (diablo fanart:0.5)
NEG: easynegative, bad_pictures, bad-hands-5, By bad artist -neg_2
35steps DPM++ 2M Karras, sxzLuma0.99VAE, clip skip 2, 512*768, CFG 5.5||

[quiet zodiac]

You could try pushing CFG up to 7.5. Maybe try "beautiful demonic woman"?

[obtuse torrent]

@slate axle ?? CFG 15 and some changes in Prompts

[rich crow]

the inpaint area is probly too small. you could try upscaling it then crop a chunk off and inpaint that if so.

[slate axle]

face seems broken kinda also

[slate axle]

what inpaint? I didnt inpaint, thats just txt2img with controlnet+depthmap

[rich crow]

ohh my mistake in that case

in that case, try inpainting a face in if you like everything else about it

[obtuse torrent]

the relative low resolution issue of the face is still true, you could try to reduce the scale of your depthmap by half, improve your prompt by adding more details about what the daemon's face looks like, and run Img2Imag together with a scale, to generate and add more details in the process... .....
you may not get great results with a promp that is too general, + guide process that will limit the volatility ("creativity") of the AI

[narrow dune]

im trying to generate a sky texture for my game, but there is objects in it. this is my current prompt, how can i ensure that it is nothing but a clear sky?
prompt: video game texture, blue sky, day, clouds, clear
negative prompt: distorted, unclear, blurry, nonsensical, buildings, trees, wood, objects

[obtuse torrent]

@narrow dune video game texture~~~~ ... (not sure but it can be your problem)

[narrow dune]

good idea actually

ill try without it thanks

[slate axle]

highly depends on model, res, and cfg then its rng of seeds. Is there any actual "style" you are going for? Like cartoony, semi-real, fantasy, etc.. ?

[atomic flume]

oh what model is this ?

[tidal spire]

Hey guys! has anyone tried to make a picture out of a qr code? so that it is read after? help to make such a plz))

[sudden quiver]

Anyone know what's the best way to fix messed up eyes

[silver valley]

Inpainting or by upscaling

[sudden quiver]

How do you do either of those

I'm on mobile if that makes a difference

[silver valley]

Oh okay what Programm are you using ?

[sudden quiver]

I'm just running the bot on this server

[sudden quiver]

Prompt:a long black haired man Uchiha man from Naruto Shippuden wearing cloak and mantle, highly detailed, high quality, high definition, symmetrical eyes

This is the prompt I'm working with it keeps giving messed up faces and eyes

[quiet zodiac]

Try removing any previous negative prompts and try these...

deformed iris, deformed pupils, 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, bad composite, poorly masked, nude, naked, text, words, cropped, out of frame, worst quality, low quality, jpeg artifacts, ugly

VECHNOST

[silver valley]

Yea try to use some negative prompts as Atom said

[sudden quiver]

Gotcha, thanks we'll see-after the bot stops acting weird

[misty heart]

I'm developing the Android App Voice (an OSS audiobook player).
I want to use stable diffusion to create illustrations for the onboarding and as a first step would like to create a book shelf with a few books in it.
The artwork should be in the material design style like for instance here in the screenshot or here: https://m3.material.io/styles/color/dynamic-color/imagery

However, stable diffusion doesn't really seem to have a concept of "material design" so if I prompt:
"book shelf, very few books, material design 3", the results are very weird. Does anyone have an idea, how to encaspulate the idea of "material design" so it creates useful results?

[quiet zodiac]

If you could identify any artists who who have a similar style, tag them instead of "material design 3". For instance, a desert scene in the Georgia O'keeffe style
You could also try other keywords like iconography, infographic, etc...

[stable valley]

hey there, I am sending an image to img2img to upscale 2x, and setting de noising strength to zero as I don't want the image to change, yet every time is subtly changes but enough to be recognisable. If I send to extras to 2x it, it does not change at all. Why is this?

[obtuse torrent]

@stable valley IDK the technical answer...,,,, something about "extras" only considering relative pixel positions and colors, to fill them with simple math interpolations... and Img2Img considering "new latent space" to fill based on the prompt ... or something like that 🤔 ..

[quiet zodiac]

Making some progress on my shaggy unicorn, but I still can't get the eyes to glow red.

[shell cedar]

do prompts for animations in the deforum notebook have to begin with "a"?

i want to use an init image but basically just want to impose a particular style upon it

[frozen meteor]

How do i prompt a fist in this pose, with the palm facing towards the viewer?
https://i.imgur.com/60pjsKM.png

i cannot seem to do it, i keep getting fists pointed the opposite way, with the back of the hand facing the viewer, even when inpainting over an image like this

[stiff saddle]

could anyone give me some guidance regarding models and prompts to make a realistic or semi-realistic portrait of a person?

[quiet zodiac]

Try this...

facing forward, a well lit shoulder shot of a realistic research scientist wearing a lab coat and adjusting her glasses

If you're trying to reproduce a celebrity, self or family member, submit a photo as an image guide along with your prompt. The typical aspect ratio for portraits is 4:5, but in SD speak, try an image size of 576x768.

[stiff saddle]

im trying to generate an image of Lao Tzu (an ancient chinese philosopher) but i've usually gotten anime stuff or just an old asian person that doesn't really resemble him
how do i submit a photo as an image guide?

and what model(s) is good for realistic stuff?

[quiet zodiac]

At this point, Lao Tzu is fictional/mythical. You can find images of drawings using google. Try one of those. I find this model makes realistic people quite well.
https://huggingface.co/SG161222/Realistic_Vision_V1.4

Here's an image I made using that model.

And here is Confucius produced from a drawing I found on google.

[whole magnet]

Any good loras for Blush?

realistic models

[sly gale]

Quick question as I have been exhaustively googling for the past couple weeks trying to find an answer, is there a way to override settings in prompts using automatic1111? Example: a command to set clip skip within the prompt and not the settings slider? If there is a cheatsheet for such commands, can someone pass me a link or info where to find this? If it exists.

Guess I'm up too late 😛 . If you want to reply, send me a dm? Any help would be appreciated 🙂 .

[stable valley]

At least its a thing, not just my imagination!

[silver valley]

Not in the prompt but you can add the setting to the Quicksettings then its easier to select

[compact island]

How can I use prompts?

[honest perch]

how can i get a full body image?

[quiet zodiac]

Try "head to toe". A vertical aspect ratio also helps. Yours is more portrait.

[honest perch]

what is the vertical aspec ratio config?

i have 449x449

[quiet zodiac]

Try using 576x768 or 384x512.

[honest perch]

the face looks a little bit weird :v

shit the face looks bad

[quiet zodiac]

I posted some negative prompts to help with face fix up.
#📝｜prompting-help message
Also consider trying some other models. Which one are you using?

[honest perch]

realistic vision

[quiet zodiac]

That's what I use for people. Maybe try upping your resolution to 768x1024?

[obtuse torrent]

@honest perch do you use hires.fix? generate at low resolution and do an upscaling (I.E. x2) will add better details to the generated image, (if the settings and the prompt are adequate)...

[honest perch]

how do i do that?

[honest perch]

@obtuse torrent is it any script option?

fuck off i started to have full body images and now is getting back to frames @quiet zodiac

[obtuse torrent]

@honest perch is A1111 feature

[honest perch]

Cn you do me a quick guide

[obtuse torrent]

@honest perch https://github.com/AUTOMATIC1111/stable-diffusion-webui/wiki/Features#hires-fix

un example of setting ...choose an upscaler, ...hires step = 0 ..upscaled by,,, option ( to keep the aspect ratio), Denoising strength: 0.05 - 0.55

[honest perch]

OutOfMemoryError: CUDA out of memory. Tried to allocate 9.52 GiB (GPU 0; 8.00 GiB total capacity; 2.18 GiB already allocated; 3.70 GiB free; 2.20 GiB reserved in total by PyTorch) If reserved memory is >> allocated memory try setting max_split_size_mb to avoid fragmentation. See documentation for Memory Management and PYTORCH_CUDA_ALLOC_CONF

specs: rtx 3060, 16GB ram, i5 13th generation

[sharp ember]

can i have a character lora only work on the face?

[obtuse torrent]

@sharp ember.. nameLoRA:X.x x.x is strength of lora (default 1.0) you can try x.x ---> 0.2 0.4 0.6 etc somtimes it work and a descriptive new clolthes and/or hair style.. prompt .... ie

@honest perch is a error of memory ...
the information itself does not make sense if it is not accompanied by the parameters used for the generation, consult in https://discord.com/channels/1002292111942635562/1002602742667280404...... possibly installing Tiled VAE Extension can help too ...

[honest perch]

Ok

I'm going to sleep now, but i will try tomorrow

[distant dune]

Tiled VAE is a godsend for high resolution outputs. The VRAM efficiency makes a huge difference. my favorite extension, honestly.

Yes, I am a shill.
This message is brought to you by tile gang

[cunning stag]

https://tenor.com/view/major-suchodolski-j00r-kononowicz-mojor-jajor-kujajor-gif-26498993

[candid shuttle]

Been trying inpainting to create backgrounds for character art I did by hand. Problem is inpainting keeps trying to add hair at the masking border. I'm assuming its user error, so what do I do to prevent that?

[sly gale]

I add (((hair))) to the negative prompt. Helps out more times than when I don't. Might even try leaving an outline around the character not masked and rendering that first, then adding the outline later. More steps but might help 🙂 .

[grizzled steppe]

any tips on how to get this ai

cant find it anywhere on the internet

[frigid iron]

Quick question, does applying e.g. :1.4 at the end of a bracketed collection of prompts e.g. (abc, aba, cd:1.4), apply 1.4 across all the tags within the bracket or only the final one (cd)?

[summer robin]

Hi, I am a newbie of Stable Diffusion.
I'm using ML-Stable-Diffusion on Mac M2 tip with python.

Anyway, I have a question.

I want to remove text from the images generated by Stable-Diffusion.
I feel negative propmt doesn't work well.

Could u help me ?

I specify like below

prompt = "lifestyle"
negative_prompt="text" 

but there always are texts on the image.

I also tried like the way below

prompt="lifestyle"
negative_prompt="text, signature, watermark, username, artist name, stamp, title, subtitle, date, footer, header"

Could someone pls give me an advice?

This is an example

[dull glacier]

Hi guys, I'm a bit new to using stability.ai and I'm looking for modifying my existing image based on the prompt. But it seems to be modifying the image entirely. Don't know what to do. Please help.
Following the documentation here::
https://platform.stability.ai/docs/features/text-to-image?ref=blog.streamlit.io

Example: I have an image of a house covered in snow. I want to turn that image so that one can see how the house might look when it's bright and sunny.

[quiet zodiac]

@summer robin Your negatives look good (perhaps add "words,font"). Try using a different model. Sometimes, nothing will get rid the words other than a new random seed. Your prompt is kind of vague, consider adding to it.

@dull glacier It might be easier to start with an image of house without snow as a generic profile, then generate one with snow covered roofs and one on a sunny day.

@grizzled steppe That just looks like a series of image guides (video footage turned into an image sequence) run using a fixed random seed.

[wanton crypt]

So this program works like this: You type in what you want to draw and AI does the picture. Right?

[honest perch]

yeah

[honest perch]

example of an image i did

[wanton crypt]

Ok. I don't know how much my current machine can process pictures. It has only 8GB of RAM.

[silver valley]

That should work, more important is your gpu and vram

[wanton crypt]

My GPU has 6GB.

[silver valley]

Okay than you can use SD localy on your PC

[honest perch]

install v1

[wanton crypt]

Thanks for info. I try this someday.