📝｜prompting-help

[tired vigil]

ok

thanks

[honest granite]

Hello Everyone! I'm very new to this and would love some direction for ControlNet and Deforum prompting please! Thanks in advance!

The situation: I want StableDiffusion to follow my video inputs more closely to the prompts I give.

My Process: I've gotten ok results using ControlNet's preprocessors and model counterparts for Canny, OpenPoseFull, Depth Leres++, and LineArt to the point where the outputs can follow the video, but the "animation" doesn't really follow my prompts that well. Such as "a soccer ball made of fire, flaming sphere, conflagration, fireball, smoke trail, retrofuturism" to get a video of a soccer ball to look like it's an illustrated flaming ball of fire. The result is not even close to what I'm thinking. I might not be using the prompts correctly and would appreciate any help! I've also tried different samplers such as Euler a and DPM++2M Karras with no real differences.

Caveats: I noticed when I switch the Animation Mode, under the Keyframes tab in Deforum, from Video to 3D (while also having all the relevant paths to the reference video in each of the ControlNet model tabs, as well and the Init tab), it no longer follows the video but just repeats the 1st frame. What it outputs in 3D mode is much closer to my prompts, it just doesn't follow the video!

Questions: Why is there a disconnect between video input and 3D animation mode in the keyframes tab? Do I need more/less detailed prompts? How can I improve the weight of the animation so it follows my prompts "correctly"?

More Info: I've attached a few settings files, and a sample of their corresponding output frames, just in case someone would like to take a look and see what I'm doing wrong. I can't thank you enough! I'm using the following stuff -
RevAnimated v122 checkpoint: https://civitai.com/models/7371/rev-animated
Deforum: https://github.com/deforum-art/deforum-stable-diffusion
ControlNet: https://github.com/AUTOMATIC1111/stable-diffusion-webui/discussions/7784

[obtuse torrent]

first copy the metadata to the clipboard by clicking on the first image....@tired vigil
at the blank prompt you copy the contents of the clipboard .... ctrl+V ,,,
then the icon of the third image is pressed .... finally you have to manually change the things that are not copied automatically or that you do not have installed (model, upscalers, extensions used)

[tired vigil]

i cant find the metadata

where do i find it?

[frosty trellis]

How do I get this robot arm connected to her back on this prompt?

1girl, full body portrait,eye level:-1), beautiful, very pretty, extremely beautiful, bright blue eyes,(short hair, grey hair,punk rock hair:1),ear rings, perfect light skin, sexy,crop top,cleavage, (large perfect breasts), perfect hands,(exposed stomach, exposed belly:1.2), (abs:0.9),high thong, thick thighs,big ass,league of legends style, cyberpunk city, streets, night, modern city, skyscrapers, moon,lora:PerfectFullBreasts-fCV3.5:0.5,lora:more_details:1, exosuit, machinery, military, yellow upper armor,shoulder armor, yellow mechanical arms, solo, standing, looking at viewer, pants, yellow boots,additional mechanical arm, lora:nijiarmor_v2:.08
Negative prompt: (worst quality, low quality:1.4), By bad artist -neg, bad-hands-5, bad-image-v2-39000, badhandv4, verybadimagenegative_v1.3, EasyNegative,(ugly, jpeg, jpeg artifacts:2.0), ng-bad_artist, ng-bad_artist_anime, ng-bad_dream, ng-beyond_negative-v1.2, ng-deepnegative-v1.75t, ng-easynegative-v2, ng-new_negative-v1.4, ng-unspeakable-horrors-64v, (worst quality, low quality, lowres, poorly drawn:2.0), out of frame, (extra fingers, deformed hands, polydactyl:1.5),asian,leather, vinyl
Steps: 64, Sampler: DPM++ 2M SDE Karras, CFG scale: 6.5, Seed: 637054513, Face restoration: CodeFormer, Size: 512x664, Model hash: 4199bcdd14, Model: revAnimated_v122, Denoising strength: 0.71, Hires upscale: 2, Hires upscaler: Latent (antialiased), Lora hashes: "PerfectFullBreasts-fCV3.5: 92a2a4b058fe, more_details: 3b8aa1d351ef, nijiarmor_v2: d1594b2c203d", AddNet Enabled: True, AddNet Module 1: LoRA, AddNet Model 1: nijiarmor_v2(d1594b2c203d), AddNet Weight A 1: 0.7, AddNet Weight B 1: 0.7, AddNet Module 2: LoRA, AddNet Model 2: more_details(3b8aa1d351ef), AddNet Weight A 2: 0.85, AddNet Weight B 2: 0.85, TI hashes: "bad-hands-5: aa7651be154c, bad-image-v2-39000: 5b9281d7c6ad, badhandv4: 5e40d722fc3d, verybadimagenegative_v1.3: d70463f87042, easynegative: c74b4e810b03", Version: v1.5.1

[obtuse torrent]

@tired vigil.... in civitai you click the image ans it open a new one ...bottom right is the button ...

[tired vigil]

Oh thanks

I'll do this tomorrow

(because me and my big brother share PC's)

[urban void]

Need help,how to do any text woth fur coated ?

[frosty trellis]

Go to civitai and search there .. you should find some good ones.

also search this discord for fur coated and I saw several prompts show up

[cloud marlin]

hello so i want to create 2d effect for my game like explosion fire etc, with the ai but it just gives me like fireing houses anyone knows what i need to write to work and give me fx?

like this

[teal lark]

Could someone explain why all the images I generate look really terrible?

Im using anything-v5

[silver valley]

you need to use a vae

that will fix the colors

[urban void]

Need more CFG scale(automatik 111)/Guidance scale(easy diffusion ) parameter,7+...

[frosty trellis]

#📝｜prompting-help message if anyone has any ideas i would be so so so so greatful.

[urban void]

@frosty trellis try to use KUKA ARM

https://www.google.com/search?q=how+to+add+object+to+picture+online&newwindow=1&ei=ma3NZIqkN5mrwPAP5YO74AI&oq=how+to+add+object+to+picture+online&gs_lp=EhNtb2JpbGUtZ3dzLXdpei1zZXJwIiNob3cgdG8gYWRkIG9iamVjdCB0byBwaWN0dXJlIG9ubGluZTIIECEYFhgeGB0yCBAhGBYYHhgdMggQIRgWGB4YHUigmxVQjbMTWNKLFXAEeAGQAQCYAYkCoAHQLqoBBjAuMzIuNbgBA8gBAPgBAagCAMICChAAGEcY1gQYsAPCAgoQABiKBRiwAxhDwgIFEAAYgATCAgsQLhiABBjHARjRA8ICBRAuGIAEwgIHEAAYigUYQ8ICBxAAGA0YgATCAgcQIRigARgKwgIGEAAYFhgewgIFECEYoAHCAgQQIRgV4gMEGAAgQYgGAZAGCQ&sclient=mobile-gws-wiz-serp

[twilit arrow]

Is there a way to force two characters to appear (aside from controlnet or whatever that extension that divides the screen is)? I've tried (blah) AND (blah), I've tried BREAK ... no matter what I do, my 1boy, 1girl prompts only ever yield 1girl or 1boy.

How do you enforce what characters show up in your generated art?

[proven jasper]

latent couple extension,u basically paint an area of the canvas and tell ai you want certain thing in that area as each area gets its separate prompt

[frosty trellis]

like 1 girl standing next to 1 man kind of thing?

[frosty trellis]

I added , "Kuka arm on back" but nothing ... 😦

[frosty trellis]

I think i will just make a Lora of the character ... wish me luck

[cobalt ibex]

hi guys, i want to train a clothing style lora, the dataset i use mainly cropped on the outfit (no human face), i want to ask for the captioning should i write
"a person wearing business suit outfit ...... "
or should i just describe the outfit only like
"business suit outfit in navy color with button ....."

[cobalt ibex]

is it my webUI or the SDXL model cannot detect weight in prompts like (golden eyes:2) ?

[dawn hawk]

Change resolution to also be able to accomodate 2 characters

512x512 is very tight

[rocky idol]

a black dog --v 5.2 --style raw -

https://giphy.com/gifs/americankennelclub-akc-american-kennel-club-toy-dog-Yj2cieoL0XBEBSiKWY

https://giphy.com/gifs/flawless-belle-gugu-mbatha-raw-e4ankQjCulbYk

[high walrus]

the bottom sailes were inpainted in a attempt to remove them, but it only made them see through

[solar dust]

Which prompts do I use to make a character posing like this?

I've tried battle stance, kicking, roundhouse kick, one leg up, etc but it never reaches that pose.

[tired vigil]

sounds like a job for controlnet openpose or training a custom lora/checkpoint

[solar dust]

Wth is controlnet openpose?

[tired vigil]

you can specify the pose a character is rendered with that, each limb and ankle in a 3d model, easy interface but a bit cumbersome to position the joints

[solar dust]

Okay but how do I use it?

Where can I find it?

[quiet zodiac]

You could just supply that image as a guide and try your roundhouse kick prompt again.

[solar dust]

I'm not very good with using img2img.

[tired vigil]

help

where do i find clickskip?

[worldly shadow]

How do I make stable diffusion not do something? It doesn't seem to listen to my negative prompts.

[neon crystal]

A lot of it is in the model, but you can try adding more weight to your negative prompts

[solar dust]

And how do you add more weight to it?

Cuz sometimes it doesn't listen to my negative prompts as well.

[tired vigil]

Hey I'm new here! I'd like to know how to prompt for a specific thing : If I want a person sitting at a table and on the table there should be a pair of shoes.. how do I prompt for that? It always gets messed up, trying to make the person wear the shoes or blend it with body parts and stuff. Is there a way to get around that?

[dawn hawk]

Regional prompt extension

[obtuse torrent]

A111 settings .. user interface ... + apply and reloadUI

[tired vigil]

Thanks

[bitter trout]

how can you create sprites like that? is it a lora?

[obtuse torrent]

really that's hard.. you can try taking the feet out of the scene with "Cowboy shot" + "feets out frame"... and then try to introduce "shoes on the table" and pray to RNGod... the other alternative is generation guided or Inpaint ... which add a lot of time investment

[bitter trout]

does anyone know how to fix this issue? it causes the image to be divided in two lol, its like 1/3 of my gens are like this

[obtuse torrent]

possibly you are trying to achieve an initial resolution that is too high for the model

[cloud marlin]

idk what lora means but i only type dream and the promt is a (game style what you like ) and type themed game assets or tiles, landscape,sprite.

[obtuse torrent]

@bitter trout It doesn't seem to be your case... possibly it's something related to the LOra trained in horizontal images maybe... try a slight reduction of the initial res... do a test with 480x720 (it's still a 2:3) ..and you raise the resolution even more later when you have good generations

[bitter trout]

ty

[bitter trout]

ty!

[bitter trout]

could be the case... i only get this with this model and its merges

[obtuse torrent]

@bitter trout add "solo" to positive prompt.... and "clone" ,"duplicate" to negative

[bitter trout]

ok, will try

[obtuse torrent]

@bitter trout
lol .... I thought those were part of your negatives !!... with those tags you are asking for a horizontal landscape... maybe you should consider changing your resolution to 768:512 as initial 😄

[bitter trout]

oh i was making wallpapers before this damn

forgor 💀

[fathom locust]

trying to come up with the best Lego style. what I hve so far:
positive: Lego style country field with a church. sculpture, toy bricks, centered composition, Lego blocks, symmetrical, ultra detailed, plastic
negative: sloppy, messy, grainy, ultra textured,

it's still off, any suggestions?

[next flint]

Can someone give some tips on how to not get generations to be too good? many times there is an uncanny valley feeling on my gens

[cursive cedar]

Hey everyone! I could really use the community's help. I'm wondering if Stable Diffusion SDXL can help me out with something. I usually use Midjourney, but I'm not the best at diffusion models. Here's what's going on: I run an online fashion e-commerce print-on-demand store. I rely heavily on mock-up images from POD sites like Printful, but the quality isn't great. When I create mockup images with real people wearing my products, they often look boring and not very good. Unfortunately, I can't afford to do any photoshoots, and I haven't been able to find a good mock-up generator that works well for complex products like dresses. So, I'm wondering if it's possible to combine a product image (like a mock-up on a white background) with a photorealistic lifestyle image of an AI-generated person. I want this character to wear my product in different poses. Is this something that's possible with the technology available right now?

[slender pasture]

Just a simple question:

To amplify m codes I've been doing: (portrait shot:1)

But what is it to tone it down?

What if you're toning done something that has no aplifyer around it example: portrait

I want to tone down the cherry blossom petals in my code

[zealous axle]

Hi there - pretty new to SD. Am running 1.5 version on my own with the standard emaonly model.
I'm trying to generate images in a setting of the middle ages. Such as battles, keeps, troops, armored guys, etc...
However the styles / images I get in my own SD look 'qualitatively' bad when compared to what I get in Bing image creator (DALL-E) or Midjourney with identical prompts.

Bing Image Creator:

vs SD 1.5:

The ideal art style that i'd want I'm not sure yet, but so the standard I get from SD (or through art style prompting) sadly do not satisfy. Here I specified 'art style of middle ages', where SD is arguably more realistic, but when leaving it out, it doesn't make a lot of difference and I do like the DALL-E style more. The https://www.midlibrary.io/styles/medieval-art Medieval art - v4 is also one I really like, but unsure how I should be generating that.

I tried to lookup some info regarding art styles in the hopes that would clarify, but engineering art styles in the prompt wasn't really fruitful.
Any hints or pointers? Should I be using different models, ... ?

[high walrus]

you will probably need to inpaint extensively its hard to make a full on battle scene

[solar dust]

So you're asking for your results to be bad?

Well that's easy, just don't prompt, upscale or use models/vaes properly.

[vivid violet]

does anyone know if negative prompting in the positive prompt works in A1111, I can't find any clarity on this. For instance prompting (hat:-1.0) or (hat:0.0) or something like that to indicate that you don't want that in the positive prompt

[quiet zodiac]

You could try the standard Stable Diffusion de-emphasis symbols know as brackets. Parenthesis boost, brackets reduce weighting. Otherwise, if you don't want the term in your prompt, remove it.
[hat:1.5]
I would think a positive boost of 0.01 would result in a reduction of weighting, however.
(hat:0.01)
I'm not sure if weight values less than 1.0 are allowed.

[prisma grail]

How to generate icons and objects? most of my models fail at most prompts

[prisma grail]

hello?

[neon crystal]

you can add parentheses like this: (negative prompt) or number like this: negative prompt:2

[prisma grail]

I believe (negative prompt:1.x) above 2 might cause problems

[neon crystal]

depends on the model and the prompt you add weight to

experiment to get the best results for you

[prisma grail]

yes

which is prefered?
hourglass
hour_glass
hour-glass
hour glass

should I use space?

I feel like it sometimes splits up the words as tags

[hard elm]

What are you trying to make, and what model are you using?

[prisma grail]

I have tried using anything v4.5 and revAnimated however it's not very efficient...

I tried using the "gameIconInstitute" model but it didn't seem great when it came to variety

[hard elm]

Have you used the bot at all?

[prisma grail]

I have not?

What bot-

[hollow tapir]

There shouldn't be any "better" one. Just different results. Words are converted into a mathematical values / token anyways. Sometimes a word can even be split into multiple tokens. You'll probably get more predictable results by sticking to the most commonly used syntax as it will match with more labels/tokens from the training set.

[prisma grail]

hmm... I see

[granite crescent]

is there a way to make the AI understand a color name is related to color alone? For example, if i prompt forest green shirt it will literally put a forest in the background.

[stable turtle]

I think you can say like color:{forset green shirt} maybe

[granite crescent]

i will try this now, thanks.

[gaunt geyser]

Hello everyone,
It is almost impossible to achieve realistic colors and proportions to create a national flag... Does anyone have dealt with similar challenge when it comes to asking for an abstract concept such as country flags?
Here is an example of the the promt "german flag" (sdxl 1.0)

[gaunt geyser]

yeah that should work better you can specify further "shirt:color:forest green:1.5"

[stable turtle]

I belive the best way for prompts is to have an LLM write the prompts for you while you give all the descriptions and colors of what you want in an image.

[hard elm]

#1100170312106127410 Any of the bot channels for SDXL; when it comes to prompting, I would start small. Keep it simple

And build your prompt

[solar dust]

https://github.com/hako-mikan/sd-webui-regional-prompter

Try using this, should help out once you learn to use it.

[gaunt geyser]

Hey thanks for your answer. Don't know what LLM is ?

[stable turtle]

large language model chatgpt for example

[solar dust]

Does anyone know a model that can generate images with the artstyle of Invincible?

Or at least similar to?

[gaunt geyser]

I would suggest find the artist(s), that is the important part I think. I don't recall a specific model for that kind of style but I'm sure you can find a LoRA for it.

[lavish trench]

I still wondering if breaking lines is a good technique... 🤔
What do you think?

[gaunt geyser]

This has potential. Thanks a lot!

[solar dust]

You're welcome.

[crude oracle]

hope i may ask since new. is this extension i need to install?

the regional prompter i mean sorry

[obtuse torrent]

----------------How to write the prompt? ...It all depends on the model, IE:

1.- How was the model trained (the model is a fine tuning) .... what was used to train it: the initial huge data set? just a part of it? a brand new one?
2 What processes and clip does the model use to identify the tokens and then vectorize them and communicate with the Unet.
Everything can have or lose importance depending on the context in which it is applied... signs "()[]{}"":,_ -"
use tags, use common language, the UIs are changing, they are being altered by scripts or extensions continuously

[orchid ore]

@solar dust would try westernanimation model on civitai, but not sure it can do invincible, not expert on superheroes.

[solar dust]

Idc about the superhero stuff tbh, I just want the drawing.

[crude oracle]

i use realistic vision V4 or 5 mostly

[orchid ore]

https://civitai.com/models/86546/western-animation-diffusion

[crude oracle]

thks

[solar dust]

yw

[obtuse torrent]

@gaunt geyser for models based in tag (Booru normally) ..a whitespace or comma separator will suffice..."-" and "_underscores" shouldn't be a problem, but newer "common language" based models don't handle them well ..
LLM thing is an AI helping you to reduce your texts to a short paragraph that is easy to understand so it is oriented to normal writing

[frosty trellis]

I typically add it as a prompt and it styles it that way. "Invinsible cartoon style" or something like that.

Any advice on how to create a lora that will give me extra kuka like appendages? #📝｜prompting-help message

[solar dust]

I just use Loras, no idea how to create them.

[toxic saffron]

anyone know a good tag / word for fangs / vampire teeth? i can't get sdxl to produce them :( tried other sdxl models too, without success

[frosty trellis]

Anyone have any advice on how to create a lora that will give me extra kuka/robotic arm like appendage? Like in this pic.

[granite crescent]

by the way what does the number 1.5 do here? does it target a certain level of the image?

because i prompted shirt:color:blue:1.5 and pants:color:red:1.5 and they got mixed up

[gaunt geyser]

in use of text:1.5 1.5 adds the weight of %150 instead of %100. It does the same job that brackets do as in (text)... Its simply more practical to use 1.5 instead of (((((text)))))

[granite crescent]

oh ok thanks.

[gaunt geyser]

If I may, I would recommend using segmentation in controlnet for such a job.

[frosty trellis]

Here is the Lora for it https://civitai.com/models/49968

[frosty trellis]

Depending on where that prompt comes from. most prompts only are -1 to 1. some do go from -2 to 2 though.

[frosty trellis]

Can you still use the "+" or "-" as well?

[clever shore]

Are the prompts that the bots add for the styles known? I want to get them on my own local sdxl

[ionic vault]

can I use any of the bots to generate image to image. or use any ref image? I couldn’t find any info about that. if so, how to prompt ? please

[frosty trellis]

https://youtu.be/ddcySSbjkNQ

[scarlet night]

Can someone help me to create tileable images in DreamStudio? Any prompt advice for repeating patterns?

Or even how to change the color pallet? I am getting amazing results, but they are all the same colors as my source image.

[scarlet night]

@scarlet night sees tumbleweed roll by as a haunting whistle is heard in the background

[amber fable]

he peepz, i have an image of 2 words in golden color. i want to give it more structure and so on.
does anyone have some prompt for me to maintain the words and style?

[frosty trellis]

whats your prompt?

Still having problems getting robotic arm to show. I used img2img of the image below and was able to get the second image. I was so happy but I cant get it to take any other poses in img2img. I tried to use control net in txt2img using the second image with the same prompt as a reference from img2img and it will not draw the robot arm in txt2img. I wonder if any of my negative prompts are stopping it? Any help would be appreciated. I think I am going to have to create a Lora with images of the original??

Prompt:
1woman, solo, (((robot arm on shoulder))), hardsuit lime green,upper armor, nice hands, perfect hands,beautiful face, perfect face, full lips with (grey punk rock hair:1.3), earrings and a (headband lime green),(perfect breasts:1.3),huge breasts,tattoos, wide hips, abs:1.5,,thong,lace,garter belt,(bluejean shorts:1), mechanical legs, sexy:1.5,(4K, unreal engine,perfect shading, atmospheric, ultra realistic, dark colors + complimentary color highlights + hyper realistic, volumetric light, cinematic lighting, cinematic detail, perfect color grading, light and volume, cinematic lights, golden ratio, studio lighting, rim lighting, photorealistic, soft shadows, 8K, reflective and refractive gleam sheen, glistening elements, edge lighting, raytrace, volumetric light, maximum detail, maximum texture, (face in focus),<lora:nijiarmor_anime_v2:0.45><lyco:GoodHands-beta2:1><lora:GoodHands-beta2:1><lora:mechaarmor_v0.7:0.15><lora:PerfectFullBreasts-fCV3.5:0.55> Negative prompt: (worst quality, low quality, poor quality, bad quality:1.35), EasyNegative, (monochrome:1.1), (greyscale),(ugly, jpeg, jpeg artifacts:2.0), (worst quality, low quality, lowres, poorly drawn:2.0), out of frame, (extra fingers, deformed hands, polydactyl:1.5),leather, vinyl,skinny,grainy,thin,kids,kids face, gun,holding,rubber, stretch,stretchy,spikes,spike,snake,rubber like,tube Steps: 46, Sampler: DPM++ 2M SDE Karras, CFG scale: 7, Seed: 1537486255, Face restoration: CodeFormer, Size: 518x720, Model hash: 4199bcdd14, Model: revAnimated_v122, Denoising strength: 0.7, ControlNet 0: "preprocessor: none, model: None, weight: 2, starting/ending: (1, 1), resize mode: Just Resize, pixel perfect: True, control mode: ControlNet is more important, preprocessor params: (-1, 0.5, -1)", Hires upscale: 2, Hires upscaler: Latent, Lora hashes: "nijiarmor_anime_v2: d1594b2c203d, GoodHands-beta2: c05ed279295e, mechaarmor_v0.7: 09ace1a26212, PerfectFullBreasts-fCV3.5: 92a2a4b058fe", TI hashes: "easynegative: c74b4e810b03", Version: v1.5.1

[scarlet night]

Why do you have (((robot arm on shoulder))) ?? Perhaps it is confused because most things aren't described with arms on shoulders, maybe you could instead figure out how to render the arm as its own type of prompt then add that to your prompt by saying (insert the arm prompt) added to or attached to the shoulder or bicep.?

[scarlet night]

It's really like SD refuses to draw a woman with 3 arms or even a woman with an extra mechanical arm.

I sm working on a cookbook. The top and bottom pages have strips of floral patterns that extend from edge to edge. We are using William Morris inspired designs. When I request Several horizontal Strips of William Morris art in various colors, I get great results, but not horizontal

[frosty trellis]

Because that is the prompt that delivered in img

[frosty trellis]

Yes, u have to consider the negative prompt interference in abstract

[scarlet night]

These would be awesome if the flowers face up when on their sides.

[frosty trellis]

It actually comes from the shoulder plates and is from the center of the back. That would need a Lora.

[frosty trellis]

Try and describe extra hydraulic arm that extends from back of shoulder armor/plates?

[frosty trellis]

I'm all ears. I could only gthat would be awesome.. I'm all open to a suggestion prompt for that

[frosty trellis]

I tried 3rd arm and got some crazy results. I have been using revAnimated_v122 and a few others but rev is the closest I have found.

[scarlet night]

Anybody know of a way to generate seamless tile patterns ?

[orchid ore]

in A1111 there is switch tile for it.

[eager compass]

how do I pass image ref

[frosty trellis]

Hey WV, thanks for replying. I have tried that. You mean like 1woman, (robot arm param here). How would you prompt for such a thing? I have tried so many different prompts with minimul prompt arguments and still nothing. The problem has been when I use non "arm" terms nothing happens. Using 'arm' has been the only way to produce something at all. 😦 I was thinking of making a Kuka Lora OR do a Lora of the images I have made or have with the robot arm and reference that. I tried a Lora of that creation but it didnt go well. I got the base image done using img2img. My prompting skills are lacking I am sure of it. How would you prompt for something like that coming from the back. The robot are is actually attached to the back of the shoulder plates of the character. I tried that but got crazy results. Some of the prompts I have used to no avail 1woman, solo,with a hydraulic arm attached to her shoulder with two mechanical arms by her side,1woman, solo,(((yellow hydraulic arm))),(((hydraulic arm coming from shoulder))) - this one produced a very malformed robot arm but was probably the closes to getting something without using img2img, (((yellow hydraulic arm))) - malformed results but something ... How else would you describe a robot arm on someone's back to an AI without using arm? appendage? limb? I have been working on this for days. 😦 I lost it when I got img2img to produce something. I think I danced around the house... LOL

[frosty trellis]

I found a Lora for multiple human arms ... that was a disaster LOL

Also the negative prompts can take out extra limbs so you have to be aware of whats in the negative loras/prompts.

[frosty trellis]

Like tile of poses or tiles of design elements like rows of patterns like paisley or plaid? I found this https://www.howtogeek.com/832164/this-stable-diffusion-fork-can-generate-tiled-images/

[frosty trellis]

You mean like using it as a base model for your prompts? Try img2img or controlnet.

[frosty trellis]

This?

[frosty trellis]

So the side of the book is great but you want the flowers on the bottom as well. I wonder if you can just apply the entire image as a mask for the book or are you trying for just one of the flower rows?

When I run the image through Interrogate CLIP (terrible wording there devs) I got this a woman in a futuristic suit with a gun in her hand and a sci - fi fi fi fi fi fi fi fi fi fi fi fi fi fi fi fi fi fi fi fi fi fi fi fi fi fi fi fi, Eve Ryder, biopunk, concept art, shock art LUL

no arm

[harsh hamlet]

Does anyone know how to generate an art like this? I'm pretty new to Stable Diffusion (Just learned it yesterday, using A1111's DirectML Web UI) and I'm trying to generate an image that contains myself standing infront of either a fighter jet or a helicopter (Just like the images I sent). I understand that I have to set up Dreambooth, download (and maybe merge?) different models/LoRas, experiment with the prompts and settings, and maybe train the models I used by using a dataset but all the fighter jets i generated just look kinda off? (displaced landing gears, fused proportions, etc). I tried "RAW Photo, Tom Cruise standing infront of F16 fighter jet, wearing a tuxedo, 8K UHD, DSLR, soft lighting, high quality, film grain, masterpiece, ultra high res, Fujifilm XT3, (photorealistic:1.4), detailed textures, tack sharp, cinematic, highly detailed, colorful, modern photograph, fighter jet, airplane, aircraft, sunset, military, dof" and have used the "Realistic Vision V5.1" and "MajicMIX Realistic" models but I'm nowhere as close to these photos. If somebody can school me on this, I would gladly appreciate it.

[turbid vapor]

any SD XL replicate users? wondering if there's a way to get weights/arguments {i.e. (((prompt here: 1.542)))} to work when using that platform

[tired vigil]

HI, I want to create an image where someone holds an (cabled) telephone handheld and a fist comes out of the handheld and smacks this man, with a POW! sign as in the old Batman comics... but I cannot make DS understand me ... I even tried ChatGTP to create a prompt, but ... nada ... all I get is this

or this "fist comes through telephone handset striking head of person holding the handset, with "POW!" as in Batman comics"

[frosty trellis]

I used this to do the same for a character i was using. https://www.youtube.com/watch?v=QLLfxQ6JoLQ Masking is the key.

[frosty trellis]

I saw this and thought it might help you. https://civitai.com/models/125194

[solar dust]

What's the model you used for this one?

[orchid ore]

@frosty trellis yes this, but probably difficult to get some good result. Just this is for it

[tired vigil]

thx tacoturtle...! am not sure how to sue this... do I copy that image and upload it as an image as "basis"? thank you

[tired vigil]

Thx Rye... not sure what a model is in this context... is "Batman" a model then? Or is that images I upload? sorry and thank you

[frosty trellis]

what are you using here?

Side Note: This prompt is useful if you are having issues getting things to color properly. (headband:color:lime green)

[tired vigil]

Is it SDXL v1.0 then?

[frosty trellis]

What did you pick in that drop down? thats the model.

[tired vigil]

thx!

[granite crescent]

when you are trying to generate a complex character with a detailed costume, is inpainting the only way to get everything you want? regional prompting got me 70% of the way there but still missing that 30% of character detail.

[frosty trellis]

I would look into taking the original model and editing it so that its parts you need like the hand slap and then use img2img controlnet reference to start

[prisma grail]

what's up with these fingers? I'm using bad-hands-5 and to what I remember I never have had this poor of a result?

is it enough for hands or do I need something?

[solar dust]

bad-hands-5 also doesn't seem to work that well for me.

[quiet zodiac]

Perhaps try more iterations? That can sometimes fix up bad fingers.

[frosty trellis]

https://civitai.com/models/16993/badhandv4-animeillustdiffusion https://civitai.com/models/47085/envybetterhands-locon negative prompt (extra fingers, deformed hands, polydactyl:1.5)

[prisma grail]

okay

[rigid radish]

The little nuances I used to use in SD 1.5 for interesting poses and angles seem to have no affect in XL and it's driving me crazy.

[leaden socket]

Anyone know a good way to get items? I'm trying to get a stone knife thumbnail for a game item.

[leaden socket]

These are still great.

[sinful oar]

Badhands helps on 1.5 models, but by no means is it fool proof. Either generate more and pick the best, or fix it with inpaint+controlnet

[frosty aurora]

Does anyone know if there's another way to use wildcards, but without the random. using | from left to right?

[dusk wagon]

Hum, did you really have 'this poor of a result' ? I don't remember this lora having like 100%, far from it with hands

[leaden socket]

These are perfect there's just multiple of them 😄

[frosty trellis]

What was your prompt?

[hushed cove]

Any idea on how to make it more realistic but still cute ? A baby panda attempting to climb a tree, looking unsure, highly detailed, professional photoshoot, fog around, volumetric light, amazing crisp 8K detail no noise, DOF

[leaden socket]

one primitive ancient flint knife ui item, realistic asset item, full view, single item only, product shot

[quiet zodiac]

Try using a different model? This is Realistic Vision 1.4.

[tired vigil]

Do you guys know how to remove these lines ? I get these when using controlnet / sd upscaler

[pastel juniper]

Hi. 🙂 I wanted to try something new and haven’t really found anything similar to what I’m looking for. I am trying to replicate the style of the game Cuphead. That vintage cartoon 2d art. I keep getting just anime. Anyone have any ideas on models and Lora’s for a style like it?

To clarify I want to make landscapes in that style. Not so much the characters

[glad cobalt]

hi everyone. i need help. Am trying to use the bot to create a seamless pattern. but when i text the tileability, and it doesnt match. Any idea how to achieve a real seamless pattern?

[slim basin]

Yo ! can we use a specified seed with the bot in discord ?

[quiet zodiac]

@glad cobalt MLOPS has a tileable checkbox. The other WEBUIs must have something similar.

[leaden obsidian]

this is very helpful, what I am seeing is that the comma and the parenthesis both seem to work as separators, potentially indicating multiple different subjects/objects, that said (girl) = (girl:1.1) and (blue eyes) = (blue eyes:1.1) so you had a lot of redundancy in your matrix

I haven't found one yet... did anyone reply to you?

[rich crow]

https://github.com/AUTOMATIC1111/stable-diffusion-webui/wiki/Features#prompt-editing is the best that i could find

i think that page should have most if not all on there

[next flint]

im glad it was helpful but bro that was like 4 months ago

didnt know somebody would be looking at something i posted from back then

[naive sluice]

Can someone show me an example of how to use \style\ in a prompt to force the style of a specific artist?

[next flint]

have you first checked for a lora or ti of their style

[naive sluice]

Haven't learned how to use Loras or tis yet. Wanted to start with the basics.

[next flint]

which style are you looking to use

Basically what ive found works is that you put in your prompt and then you add ",by {artist/style}"

[naive sluice]

Experimenting with Chagall.

[naive sluice]

Including the curly brackets?

[next flint]

yeah. they are like an updated version of parentheses that are less likely to accidentally "eat characters" but u can use parentheses as well

[quiet zodiac]

You can simply type..."in the style of my artist" at the end of your prompt.

[solar dust]

Assuming the model knows said artist.

[slender pasture]

idk if this counts as "prompting help" but im having issues with quality in img2img?

🤔

Idk what to ask really, I guess general direction on quality improvement strategies?

[proven jasper]

increase the resolution

[slender pasture]

🤔 change the heightxwidth?

[proven jasper]

yes as high as your gpu can handle

[slender pasture]

well i pushed it to 1080 x 1770 with the lil button (that i didnt see) that makes the output the same size as the input 🤔

Ill figure it out from here- ty

[proven jasper]

highest img2img can go is 2048 x 2048 i think

[slender pasture]

well I cant run it at 1080, but 720 looks a bit better-

still fucky but its improved a lot, so thank you for pointing me in the right direction 🙂

[proven jasper]

you can go higher regardless of gpu with the tiled vae extension

[slender pasture]

🤔

I know a VAE is that thing that goes in the settings 🤔

[proven jasper]

also you can go beyond 2048 px resolution with another extension called ultimate sd upscale

u can combine both and get an almost perfect 4k image

[slender pasture]

Interesting, wish I understood the technicality of all this more 🙃

[solar dust]

Try using an upscaler

also revise your negative prompts.

And make sure you're not using a broken upscaler

Anime6B has been know to not work properly for quite a while now.

[next flint]

thats the highest the default settings in the webui allow. u can get around it by either putting an image in and using the button to autoget the dimesions or by changing your webui settings to allow higher values

[next flint]

Do you still need help w this?

[next flint]

If you're using SD upscale id recommend switching to Ultimate SD upscale because there is a setting called "Seams fix" which helps get rid of those

[naive sluice]

Thanks LC

[naive sluice]

If you are using SDXL you need to use 1024x1024 then upscale. It was trained on 1024.

[tired vigil]

I'll try that later and yes I was using the ultimate one didn't know a non ultimate was a thing :b

Thanks

[fallow ginkgo]

Hi I was working with SDXL but it stopped working I have one laptop and one desktop one of them is with dark blue background and the other one is with white background... They are both stop working ... After I generate it did not give me result

[balmy kernel]

What makes for a good generic Negative Prompt for SDXL?

[quiet zodiac]

Try leaving the negative prompt blank until you actually need it. Don't include the word "no" in your negative or you might introduce a double negative...no bad hands for example might give you bad hands. Also avoid the word no in your positive prompt. Tell SD what you want, not what you don't want.

[sour notch]

Hello friends
I have a problem with the text
to image
I want to have different clothes in the picture of a man and a woman, but I don't know how to distinguish a man's clothes from a woman's.
For example, I want the man to wear a black hoodie and the woman to wear a red t-shirt, but after generate, both of them are wearing a black hoodie.
How do I write prompt?

[solar dust]

It's very hard for AI to make 2 separate people each with different clothing but

try writing it like this

1girl (red t-shirt) and 1boy (black hoodie)

you might not get it right at first so keep trying.

[next flint]

there is something called regional prompter

ill send the github link in a sec but basically it lets you divide the image into different areas and you can prompt accordingly

https://github.com/hako-mikan/sd-webui-regional-prompter

[solar dust]

Wouldn't that be a problem if he wants the couple in the image to interact with each other?

Like holding hands, kissing, etc?

[next flint]

i dont think so

[sour notch]

🙏

[next flint]

like look how much this was broken up

[solar dust]

Yeah but that is one person in the image.

[next flint]

well firstly it depends on if they are using sdxl or v1.5

[sour notch]

tnx bro i see now

[next flint]

because controlnet + reigonal prompter would easily be enough

since controlnet sdxl is yet avaliable on all webuis is hard to say

[sour notch]

I want exactly
To illustrate a music video

[solar dust]

@next flint Is this the one?

[next flint]

yeah i believe so

also tiled vae has it built in

[sour notch]

I think it is possible to extract positions from real photos and use them in your controlnet

[next flint]

yeah you can either extract a rough skeletal stucture or just the outline

[sour notch]

Thank you very much for your time Bros👊

[sour notch]

This article is interesting and useful, thank you for introducing it

controlnet solved it to some extent as a test

[solar dust]

You already learned how to use it?

I'm finding it quite complicated.

[next flint]

I think control net has become a lot more accessible

After 1.1 things became easier to use and the webui extension became more clear about what things do

[next flint]

could you send the image you used in controlnet

[next flint]

Im not sure if this is what your looking for but i was able to get this after toying with it for a bit. Could definetely be improved with a more specific prompt and adetailer

[sour notch]

lord How did you show the colors of the clothes separately to the artificial intelligence?

[next flint]

i use regional prompting and sdxl

i just decided to go with a real photo style since i have a set of tokens that works really well for it

if u put one of those images into a site like jimpl it will list the entire workflow

Like if you set the sections right you can get stuff like this. Im really new to even trying regional prompting but it seems to work really well for specifying things

[sour notch]

png info tab
gave me all your details
I used all your comment and your photo
And it gave results
The point of the work was in the prompts that you wrote in a specialized and expressive manner
tnx bro

[next flint]

tbh the prompts i use are mostly just from my gpt prompt gen lol

but im glad i could help

[sour notch]

nice

Despite some minor errors in the image (such as fingers), the result was interesting considering your settings and reference photo

[next flint]

I doubt it would help much but i would recommend just going for a giant negative prompt

[leaden obsidian]

@rich crow Trying to figure out how nesting work with commas and parenthesis. I have noticed that despite people saying it isn't important it actually is. As your picture showed @next flint the placement of commas often leads to the model interpreting the word as a seperate occurance of the object. However I can't find a detailed explanation on how this works anywhere. Thus I found myself searching back many months looking for people talking about commas.

[next flint]

honestly ive found that {} work better.

"On 2022-09-29, a new implementation was added that supports escape characters and numerical weights. A downside of the new implementation is that the old one was not perfect and sometimes ate characters: "a (((farm))), daytime", for example, would become "a farm daytime" without the comma. This behavior is not shared by the new implementation which preserves all text correctly, and this means that your saved seeds may produce different pictures. For now, there is an option in settings to use the old implementation.

NAI uses my implementation from before 2022-09-29, except they have 1.05 as the multiplier and use {} instead of (). So the conversion applies:

their {word} = our (word:1.05)
their {{word}} = our (word:1.1025)
their [word] = our (word:0.952) (0.952 = 1/1.05)
their [[word]] = our (word:0.907) (0.907 = 1/1.05/1.05)

"

[next flint]

The previous grid i made is kinda outdated since the models were much worse back then so i thought id make a new one and show the difference between () and {} when the prompt gets more repetitive it gets less evident but in the first few rows it leads to a difference.

i had to add a few things to the prompt and neg so that the images wouldnt get taken down but the resulting comparison is roughly the same

[next flint]

How are people getting good anime images with sdxl. In my experience with sdxl base, dreamshaper, and rundiffusion the closest i can get is hyperanime or traditional painting

[sour notch]

Can you send the sample?

[brisk karma]

/prompt

Is there any documentations somewhere that talks about all prompt commands for SDXL 1.0?

Just wondering

I'm super new to this LOL

[queen kiln]

Question about prompts Im doing anime im following danboruu tags does the tag like anime,character works? ex (dragon ball,goku) do they work even without loras? using SD 1.5

[next flint]

like for the bot or when prompting in the webui

[next flint]

like of my outputs or online

[next flint]

its hard to say. likely they wont do much unless you describe the style or use a lora/ti for it although depending on the model you might get something close to want

[next flint]

if you want to train a lora for sd 1.5 there are plenty of online options that are free (https://colab.research.google.com/github/hollowstrawberry/kohya-colab/blob/main/Lora_Trainer.ipynb#scrollTo=OglZzI_ujZq- has worked for me)

SAMIKUZA Hows the project going?

[stable gate]

Is there a quick way to generate beautiful hands with sdxl ? I am generating beautiful images with horrible hands. My negative prompt often contains "bad anatomy, deformed, ugly hands" but nothing seems to work.

[solar dust]

Try the bad-hands-5 embedding @stable gate

along with negative prompts like deformed hands, missing fingers, extra fingers, etc.

Can't guarantee they'll come out 100% absolutely flawless

https://cdn.discordapp.com/attachments/735187307103256596/1139355011860668478/image.png

but it'll be an undeniable improvement.

[solar dust]

By using the correct models and VAEs

combined with good prompts.

[next flint]

Well I understand that aspect but aside from tuned models and vae is there a way to get good results

[solar dust]

Good prompts.

And negative prompts.

[next flint]

K I’ll see what I can do

[solar dust]

You can also "cheat" if needed

by using LORAs.

If you prefer not to, perhaps try using a heavily trained model like Aniverse, which was intentionally made to generate high quality results with a low amount of prompts and no LORAs.

Aniverse is one of the best, most well-made anime models I've ever seen.

It's able to generate very high quality results with just a handful of prompts and close to no negative prompts.

[next flint]

I’m rn using a web based webui that won’t let me download any extra models so I’m looking for some tokens that would be able to at least get close to anime style

[solar dust]

You're not using a local AI? That might explain your difficulty.

I know nothing about how online AIs work so I can't help you properly.

[sweet lily]

How can I stop sd from drawing realistic teeth for anime style? I cannot stop it from happening with negative promts like realistic mouth, with

[solar dust]

Could you show me an example of these teeth?

[sweet lily]

Sure if you wanna barf

[sweet lily]

I'm using the meinapastel model.

I used the template for negative prompts in the pinned messages here.

I always ruins a good image. Is there a way to make the ai try to redo the exact image?

But only fix the mistake I point out for example inpaint?

[turbid vapor]

any SD XL replicate users? wondering if there's a way to get weights/arguments {i.e. (((prompt here: 1.542)))} to work, or if there's a replicate specific syntax? (unless they do already work -- but for me they seem not to, or at least, to have notably different output than other methods 🤔 ). if there isn't a way to do so on replicate, are there are platforms that have api functionality which would work with prompt weights/arguments? my understanding is that stability.ai doesn't allow this, either

[sweet lily]

I need help I keep getting photo realistic mouths for anime characters most of the time ruining the image. What can I do to fix it?

[solar dust]

Holy crap

that looks really bad.

I don't know what could be causing it

have you tried other models?

If they don't deliver the same teeth then the problem lies in meinapastel.

[sweet lily]

None of the preview photos have that though.

[solar dust]

Then it must be something on your end

like the prompts or VAE.

Or maybe something you installed.

[tired vigil]

You are using face fix aren't you

Don't use that garbage utility for anime,Pixar... Or anything related to that

Guess what? Don't use it with realistic stuff either it sucks

It should be banished

[tired vigil]

Poor dude got rekt by his ui... Face fix more like fix your damn ui

[solar dust]

Don't use the "restore faces" option

that's terrible

if you want something to make better faces that ACTUALLY works, try Adetailer.

[sweet lily]

What the promt for wanting legs to be as close together as possible?

Also, why can't I get clean looking images like other people do? I seem to be doing something wrong

[next flint]

which model are you using rn

[next flint]

the site i use is just a web auto1111 ui

[sweet lily]

MeinapastelV6

[next flint]

most of the time image quality is based on neg prompt, inpainting, adetailer, and trial and error

[sweet lily]

I don't do inpainting because I don't understand it much. And I didn't even hear about adetailer. I use the negative prompt that is pinned in here as a reference. And I add some things that I don¨t want too. I also think that I have an issue with giving almost everything too much weight.

[next flint]

so inpainting isnt required but is great for fixing ai manerisms in image like warped things. the neg prompt pinned isnt a bad place to start but isnt the best idea to always use

[solar dust]

I only add excessive weight to prompts when SD strongly refuses to follow them.

[next flint]

adetailer is kinda like autoinpainting if i understand correctly

[sweet lily]

I keep looking at the Meinapastel model images on the site and I'm not getting enything close to that quality. from just generating

[solar dust]

You gotta take a look at everything

some details are not perceptible by just looking at the preview images and their prompts alone

you gotta know the VAE used, if they upscaled it, the image's resolution, how much denoising strengh and upscaling steps they used, etc

[next flint]

so this was made using a model similar to the one you are using. i just grabbed some random img parameters off civit. the one of the left used adetailer and the one on the right didnt

doesnt always give better results but most of the time it does

[solar dust]

Btw

does anyone know why img2img just keeps repeating the same image over and over?

[next flint]

@sweet lily so i downloaded one of the preview images for the pastel mix and just sent the gen params to txt2img and got this. the one on the left is from the model page and the one on the right is the one i made. i used meinamix v8 instead of pastel

[solar dust]

Please someone help.

[next flint]

wdym

[solar dust]

When I put an image in img2img

instead of it creating similar images based on the one I gave

the results are just clones of the first image.

Literally nothing changes.

[next flint]

have you tried just closing the terminal and restarting?

[solar dust]

I think I found the issue, apparently keeping the denoising at 4 or lower stops it from creating different images.

No idea why tho.

[next flint]

like 0.04?

[solar dust]

0.4.

[next flint]

i think there might be slight differences with a denoise like that but not much

[solar dust]

Yeah, I can only see noticeable changes at 7.

[next flint]

like 0.3-0.4 i normal use for small changes like fused fingers

[solar dust]

That's exactly what I'm trying to do

to fix a hand that shouldn't exist

[next flint]

inpaint?

[solar dust]

but SD just insists on letting it there.

I don't have the slightest idea on how to use inpainting.

Does it have an eraser?

So I can wipe out the extra hand?

[next flint]

like you can replace it

https://stable-diffusion-art.com/how-to-remove-a-person-with-ai-inpainting/

like if u apply this principal to a hand i think it would be fine

[solar dust]

I understood everything except for the prompt supply part

there's no option for me to type prompts in inpainting.

[solar dust]

Okay and btw @next flint

what's the one I should use?

Tried removing the hand but it's still there, just transparent now.

[solar dust]

Could someone please help me with SD Upscale?

[atomic wigeon]

Cool, but simple, prompt creator popped up today. Check it out. Safe, nothing to download or anything. I highly recommend it. He even coded it to give neg prompts! Super cool.
https://codepen.io/TheCopernicus/full/mdQZGpW

[next flint]

Ryu if i remeber correctly i think u chose fill

[atomic wigeon]

Anyone can type up a prompt. This has styles included with one click.

[next flint]

yeah sry that was kinda mean

[atomic wigeon]

👍🏻

[next flint]

ill check it out

[atomic wigeon]

All good : )

I would love to see cameras, lens and film added to it. How cool would that be?

[next flint]

but i think for that to work well for a wider audience it would have to implement a standardized picture that shows the effect of a camera term inside of the prompt so people know what they are doing

[atomic wigeon]

There is one of those already. Well, sort of.

[next flint]

this? https://huggingface.co/spaces/mattthew/SDXL-artists-browser

[atomic wigeon]

https://rikkar69.github.io/SDXL-artist-study/cameras/

Cleaner and loads faster.

[next flint]

only issue is that the prompts are different

[atomic wigeon]

I love that included the click thing too. Very nice job.

What do you mean?

This is SD Discord, not MJ.

MJ users could certainly use those tools, but weights and other stuff are different.

[next flint]

like all the images on that site use different prompts. imo (idk if everyone would want it) it would be better if they made sub catagories for artists and in the subcatagories each image had the same common prompt and then added the style

cuz then it clearly shows how it affects images

like i see that on the lenses page it does that

[atomic wigeon]

Ya all of them you just click and paste

[next flint]

but like these are all different

ig that you cant just use the same prompt for every image

[atomic wigeon]

Nope. Those are click and paste too.

[next flint]

i think we are thinking of different things

[atomic wigeon]

No one image can demonstrate every potential outcome. Filters alone can significantly differ. He's just trying to help with the prompting.

[next flint]

i get that

but still some sort of standardization to at least give an idea of what is different would be a cool feature

[atomic wigeon]

I don't know. I appreciate that the guy took he time to make it and share it will all of us, for free.

[next flint]

yeah

i agree

its nice that a good deal of these things are free

annoying when i see those sites that actually charge like 30$ for 100 prompts that are likely meh

[atomic wigeon]

That's why I shared those. Noone ever needs to do that lol.

Now we buried them under a conversation. Sharing both again. https://codepen.io/TheCopernicus/full/mdQZGpW

https://rikkar69.github.io/SDXL-artist-study/

[formal nebula]

guys IDK much about technical side of the models. do models have specific prompts also? or they are just trained on photos? im trying to figure out if prompts are general language shared in all models? what the relation between prompts and different models? I see some people use very easy text in 2 lines and get incridible good results and some other use lots of prompts like (very realistic bla bla) and still not as good a simple text. why is that?

[quiet zodiac]

Certain models do have "activation tokens" you place at the front of your prompt before they function correctly. Make sure to check out the model card. It is generally listed near the top, like this wonderful Redshift model which makes beautiful landscapes. But most don't. Some models require you use a companion LoRa file to adjust the color space. This is quite annoying, and I often avoid those. If you get a washed out or "greyed" over result you're probably missing the LoRa.
https://huggingface.co/nitrosocke/redshift-diffusion-768

[ancient wagon]

Im having trouble trying to get my prompt to constantly give me the right hair color. I'm trying to get a two toned hair that the top half is black and the bottom is white. It only occationally gives me the right color. Most of the time it gives either all white or all black.

[next flint]

can u send an example

[tired vigil]

and the prompt

[next flint]

id recommend just using regional prompter

[next flint]

so i threw the region on top of the image as an overlay to show the differences. in region 1 i added {red hair:1.25} and in 2 i added {blonde hair:1.25}

[ancient wagon]

there are regions?

[ancient wagon]

trying to recreate this in an anime style full body photo for a friend

I occationally would get something like this

But I can't get it to recreate more than once every 10-15 times

[ancient wagon]

female, Black and white ombre hair, light blue eyes,

(I use A1111 of Stable diffusion)

[proven jasper]

thats not a good prompt

[ancient wagon]

I use more than that for the prompt, thats just the part describing the head

[proven jasper]

instead of ombre hair just use (two-tone hair,multicolored hair,black and white hair:1.3)

[ancient wagon]

I tried "Black and white two-toned hair" too

[proven jasper]

black hair base,white hair ends

two-tone hair is a must

[ancient wagon]

Also tried "Black hair with white ends"

[proven jasper]

two tone hair try with that

[ancient wagon]

how would that look like as a prompt?

[next flint]

with the regional prompter extension yes

[next flint]

do you want the prompt i used or a recommendation for yours?

[ancient wagon]

either would work

I need to find the extension too

[next flint]

https://github.com/hako-mikan/sd-webui-regional-prompter

[next flint]

this is the workflow

Woman with hair transitioning from red to blonde, graceful pose, soft and natural expression, outdoor setting with lush greenery, warm sunlight filtering through trees, tranquil and serene atmosphere, high resolution, detailed and sharp focus, vibrant and realistic colors, photorealistic style, captured by a skilled photographer, professional quality, no visible imperfections, by Annie Leibovitz.,24mm, 4k textures, soft cinematic light, adobe lightroom, photolab, hdr, intricate, elegant, highly detailed, sharp focus, {cinematic look:1.4}, soothing tones, insane details, {photo, real,real shadows, real photo:1.25} BREAK BREAK {red hair:1.25} BREAK {blonde hair:1.25} BREAK

Negative prompt: {worst quality, low quality:1.4},{grayscale:1.25},{monochrome:1.25},{watermark:1.25},{white letters},{bad anatomy:1.21},{bad proportions:1.331},{disfigured:1.331},{duplicate:1.331},{low quality:2},{missing arms:1.331},{morbid:1.21},{mutilated:1.21},{ugly:1.331},2 body,2 head,2 body,2 head,2 lower,2 upper,3 feet,3 hand, abdominal stretch, acnes, age spot, aged up, bad anatomy, bad body, bad hands, bad proportion body to legs, bad proportions, blurry, briefs, cloned face, cropped, deformed, dehydrated, disfigured, dot, duplicate, error, extra arms, extra digit, extra fingers, extra legs, extra limbs, extra long leg, extra toes, fewer fingers, fused fingers, gross proportions, jpeg artifacts,long neck, low quality, lowres, malformed limbs,missing arms, missing fingers, missing legs, missing toes,morbid, mutated hands, out of frame, poorly drawn face, poorly drawn hands, signature, super long leg, too many fingers, ugly, username, watermark, weird toes, wrong feet bottom render

Steps: 20, Sampler: Euler a, CFG scale: 7, Seed: 2409681806, Size: 1024x1024, Model: rundiffusionXL_beta, RP Active: True, RP Divide mode: Matrix, RP Matrix submode: Vertical, RP Mask submode: Mask, RP Prompt submode: Prompt, RP Calc Mode: Attention, RP Ratios: "0.5,1.33,1,3", RP Base Ratios: 0.2, RP Use Base: False, RP Use Common: True, RP Use Ncommon: False, RP Change AND: False, RP LoRA Neg Te Ratios: 0, RP LoRA Neg U Ratios: 0, RP threshold: 0.4, Version: v1.5.1

[ancient wagon]

I usually keep steps around 70-90

[next flint]

ive found 20-30 is best for nearly every sampler

i mean sometime people go up to 50 but thats rare

gimme a sec and ill show it

[ancient wagon]

is there a video to show how to download this, its been a bit since ive been on github

[ancient wagon]

I've found better results on over 60

[next flint]

maybe it just depends on the model and prompt

[next flint]

either open a terminal and then cd into \stable-diffusion-webui\extensions and then clone

[proven jasper]

for a portrait 30 or 40 is good enough unless you are upscaling it to something like 4k or 8k

[next flint]

or you can use the extensions tab and paste the link

[ancient wagon]

I usually try to get near full body art

[next flint]

James Bond this is what i get with different steps

[ancient wagon]

[next flint]

like there is some difference but the time taken often isnt worth

if ur looking for better quality try adetailer (https://github.com/Bing-su/adetailer)

[ancient wagon]

I'm still getting just white or black hair instead of two-toned hair:

Prompt: Female, hair transitioning from Black to White, bikini, beach, 4k textures, soft cinematic light, hdr, highly detailed, sharp focus, {cinematic look:1.4}, soothing tones, insane details, {photo, real, real shadows, real photo:1.25} BREAK BREAK {Black hair:1.25} BREAK {White hair:1.25} BREAK, rising sun, sunbeam, light rays, full body, steam, mist,

Negative prompt: EasyNegative, (low quality, worst quality, 3d), realistic, blurry, haze, bloom,

Steps: 55, Sampler: DPM2, CFG scale: 7, Seed: 1246252195, Face restoration: CodeFormer, Size: 512x512, Model hash: c94adc132a,

[next flint]

whats the model name?

[ancient wagon]

I use a mix of 2 models

[next flint]

wdym?

[ancient wagon]

I combined 2 models from civitai into A1111 via checkpoint merger

[next flint]

i ask cuz you include in your prompt tokens indicating a real photo but youve shown examples of anime styles

what style are you going for?

[ancient wagon]

a kinda realistic with a soft touch?

[next flint]

like this?

[ancient wagon]

yea

[next flint]

alright and you said you were looking for fullbody?

[ancient wagon]

yes

or atleast upperbody

[next flint]

so then you would likely want to change your region divisions since im assuming you are using the ones i send (0.5,1.33,1,3)

[ancient wagon]

im still trying to figure out the region manager thing

[next flint]

same

[ancient wagon]

imma go to bed then mess with it a bit more tomorrow

[next flint]

k

[solar dust]

Okay, I kinda understand regional prompting more

the guide tells me to separate each prompt with the BREAK and other commands

but how do I also add a command related to 2 separate characters on the screen?

[atomic wigeon]

You don't need a "break" in SDXL.

[solar dust]

Example, if I want to generate a couple walking and holding hands

the 'walking" part can be separate but what about the handholding part?

[atomic wigeon]

Increase the width of an image to encourage multiple people in the scene.

[solar dust]

I know that already, my problem is with how to use the regional prompter extension.

The example given on the guide doesn't tell me that, apparently.

[atomic wigeon]

@next flint That 50 mile long prompt is not needed in SDXL. #📝｜prompting-help message I can ballpark your entire prompt with three words. And then dial in the details.

pose, tranquil, outdoors,

1. The word "pose" will automatically force the AI to give you a human. And most of the time it will be a woman. And most of the time it will be a glamour pose. Why? Because models pose, normal people don't. The word "pose" is very powerful and will force SDXL to give you, in one word, what you're typing it to do in your prompt. And a majority of the time the word "pose" will force a professional photograph. It's very unlikely you'll get pop art, illustration, etc.

2. The word "tranquil", when combined with pose, will trigger everything associated with the word "graceful", and far more. You'll get a lot more female specific poses. And in particular a lot of elegant and tasteful poses. The AI will merge "pose" and "tranquil" together. You'll get a lot of ballet, yoga, etc. Then dial in more specifics afterward.

3. The word "outdoors", when used with "pose" and tranquil" will force the AI to consider the conditions in which a human would be posing outdoors in a tranquil way. You will automatically get a lush green environment on a sunny day.

[solar dust]

@next flint Please, could you teach me how to use regional prompter?

I can't seem to understand it at all.

[stark sage]

Hi, is it possible to get a 3d modelling reference sheet. I tried couple prompts but it is either giving me variety of what I want or from couple angles. What I need is reference image of the 3d model from 4 sides to be able to draw it in Maya

[sweet lily]

Can someone help me correct anatomy on an exagerated anime image?

[atomic wigeon]

I'm not sure exactly what you're asking. But you could try a simple prompt like this:

Eiko Ojala style, random 3d shape,

And then adjust the seed one digit at a time.

[stark sage]

Hi thanks for your reply. When 3d modelling for example a character. I need to put 4 images to background to get reference while drawing like this:

From each side so my drawings are consistent in terms of size length

[atomic wigeon]

That's very difficult to do in SDXL 1.0. It's mostly for inspiration, not actual design.

At least not yet. Perhaps in the near future.

Companies are already working on voxelation prompting.

And microvoxels at that.

You could try to force camera changes. But it will alter your character each time.

[obtuse torrent]

if you have good reference images of what you want, human poses or well defined 3d models, consider using the controlnet extension, openpose full model... this will give you skeletons that will make the poses consistent

[next flint]

i understand that i dont need a super long prompt but what i sometimes do is that i make 2 things and combine them. i do one with natural language since sdxl is supposed to be better at that and then combine it with comma sepparated tokens for accuracy

[next flint]

like regions or something?

[next flint]

sdxl or 1.5?

[sweet lily]

1.5

too lov vram for sdxl

[next flint]

better neg

[sweet lily]

What would you add to the neg to prevent that beside everything that the pinned message in here has?

[next flint]

can you send me the prompt ur using

[sweet lily]

the neg one?

[next flint]

nah the main prompt

[sweet lily]

When i do the inpaint I do this muscular legs, best quality, correct anatomy

[next flint]

so are you getting incorrect legs?

[sweet lily]

Yes. And now most of the time they are textured for some reason when they are supposed to be just bare legs.

[next flint]

can you try this as a negative

{worst quality, low quality:1.4},{grayscale:1.25},{monochrome:1.25},{watermark:1.25},{white letters},{bad anatomy:1.21},{bad proportions:1.331},{disfigured:1.331},{duplicate:1.331},{low quality:2},{missing arms:1.331},{morbid:1.21},{mutilated:1.21},{ugly:1.331},2 body,2 head,2 body,2 head,2 lower,2 upper,3 feet,3 hand, abdominal stretch, acnes, age spot, aged up, bad anatomy, bad body, bad hands, bad proportion body to legs, bad proportions, blurry, briefs, cloned face, cropped, deformed, dehydrated, disfigured, dot, duplicate, error, extra arms, extra digit, extra fingers, extra legs, extra limbs, extra long leg, extra toes, fewer fingers, fused fingers, gross proportions, jpeg artifacts,long neck, low quality, lowres, malformed limbs,missing arms, missing fingers, missing legs, missing toes,morbid, mutated hands, out of frame, poorly drawn face, poorly drawn hands, signature, super long leg, too many fingers, ugly, username, watermark, weird toes, wrong feet bottom render

[sweet lily]

It's a little better but I'd still tear out my eyes if I ever saw someone draw it like that

[next flint]

lol

have you tried embeddings?

[sweet lily]

what that?

[next flint]

textual inversions. extremely helpful for these kinds of things

go to https://civitai.com/models and filter by these. If you use some of those it should get a lot better

[sweet lily]

Even when trying to go for super unrealistic body proportions? lige for exaple thigh muscles?

[next flint]

its been a while but i would assume it would be fine

[sweet lily]

What if I can't find the reference images that I would need? I don't understand how this workds at all

[next flint]

wdym reference images

[sweet lily]

Idk I saw some dude on yt use references for it.

So... How do I make the needed embedding for this?

[next flint]

its best to just download a negative embedding from civit

[quiet zodiac]

Reference images are very forgiving. Just frankenstien something together in photoshop that approximates your goal. It's a great way to stage the entire image.

[atomic wigeon]

Here's a fun one for you guys to try in SDXL 1.0.

Eiko Ojala style, downhill skier, centered, symmetrical, 3d,

[sweet lily]

I sketched out how the muscles should have looked like because I know anatomy quite well. Yet it just jumbled it in inpainting

[next flint]

Hadanuke this is how the negative embeddings effect an image

[sweet lily]

Ok. So in krita I painted how big I want the legs to be. Now I want the inpaint to make it actual legs. Why is that so dificult?

[next flint]

can u send the prompt

[sweet lily]

muscular legs, best quality,( correct anatomy:1.5),detailed, 8k , huge muscles, aid28

[next flint]

what about the original image

[sweet lily]

I wanted to refrain from sending it considering.... well it's not nsfw but it's not far from it either.

[next flint]

lol

you can dm me if u want

cuz if i can make the image and see whats up with it i could probably help better with fixing it

[sweet lily]

sent

[next flint]

can you send the gen data since discord wipes metadata

[sweet lily]

how to I get the gen data from that edited image?

[next flint]

if you put the image that you originally made before editing into the PNG info tab and just copy the output

[solar dust]

Like generating two people interacting with each other.

Like you did before with the images you posted here.

[next flint]

oh yeah with that samzuka guy

[solar dust]

How did you manage to make a couple hugging with regional prompt?

Did you separate it in 3 verticals and then put the interaction part on the middle?

[next flint]

these were my settings

[sweet lily]

I think I lost the original image. I didn't make a copy and I closed the editing software I used.

[next flint]

oh wait you drew that?

[solar dust]

Okay but how did you prompt the hugging part?

[sweet lily]

I drew the legs part yes.

Because it didn't want to make the legs bigger no matter what I tried.

So I instead tried this.

[sweet lily]

Nvm I got it. Used the browser extension.

[solar dust]

1girl ADDCOL
1boy ADDCOL
hugging
@next flint Is that how you did it?

[next flint]

nah it was like
Common prompt
BREAK
{a woman wearing a red shirt:1.25}
BREAK
{a man wearing a black hoodie:1.25}

ill try something in a sec and see what works

yeah i just tried it and that was basically what i did. i forgot the original prompt i used for the guy before but it was really simple. what matters is that the BREAKS split the screen. make sure that you add moderately significant emphasis to the tokens in between

[next flint]

Hadanuke sry for not responding for a bit. i looked back at the conversation and considering you are using a 1.5 model and already have and image where you have the correct sizes i would recommend controlnet

since the image could be considered slightly nsfw ill send some example images in a dm

[orchid ore]

I never used break word using regional prompt. Also have not checked anything where use common prompt is 🙂

[next flint]

For regional prompting you can use ADDCOL and ADDROW which specifically add rows but if you just want to simply break the image up once more BREAK is a simple way to do it

[orchid ore]

break is saved when saving image
@next flint 👍

[next flint]

wdym for when saving an image?

[orchid ore]

it automaticaly saving images. And ADDCOM ADDCOL and ADDROW is converted to BREAK

[solar dust]

Why { instead of ( tho?

[next flint]

sometimes ( leads to token compressing. { is supposed to be better for that. it also is 1.05 multiplier instead of 1.1

but they are for the same thing

this #📝｜prompting-help message

i mean you could use ( and likely get the same result

[silver valley]

Automatic1111 don't support {} at all for prompt weights. Like he stated in the wiki its only used by NovelAI

[next flint]

oh i thought it did

[silver valley]

Nope he switched it to ()

[next flint]

so there is no reason to use {}?

[silver valley]

Yes

[next flint]

thats good to hear. after using () for so long it felt weird switching

is the github wiki outdated then?

[silver valley]

Yea at the start of auto we uses it a lot but that was because most people copied their prompts from NAI. I have old prompts with {} too xD

[silver valley]

No why? He just explains how {} from nai is converted to () in auto1111 if some users want to switch from NAI or test their NAI prompts

[next flint]

oh i completely misunderstood it

well damn since i learned about it like 2 days ago ive been telling people the wrong thing

[silver valley]

Maybe they will figure it out if they see other prompts

[next flint]

im assuming so

[silver valley]

On civitai you won't find any {} used as weights, maybe rarely

[versed mulch]

Is it ok to use "pixiv" in bot channel?

[silver valley]

As a tag? Yes

[versed mulch]

Thankyou

[charred root]

hey guys, what words can I use to make the character return their head by some amount of degrees ?

[quiet zodiac]

I don't think prompts work that way. You could supply an image guide with the character in the pose you are trying to achive.

[atomic wigeon]

Have you ever wanted to hover a drone right over your subject content? By combining multiple views you can do just that. Here's a good prompt for ya.

woman in bikini laying on a towel extreme close up crowded beach birds eye view 3d Ricoh Theta Z1 with Built-in 14mm f-2.1

[charred root]

yep, the problem is that i don't know those specific keywords, they are probably related to photographers who shoot people, it's some professional terminology I don't know

[atomic wigeon]

I don't know about degress. But you can force the AI to make a character "looking right", looking left" looking up" looking down", etc. You can also try Dutch view. Another one is "random face pose". With enough time it will land close to what you want.

Maybe try "looking at 2 o'clock high"? Never tried that but it might jsut work.

[charred root]

hmm, interesting, I'll try that, thanks @atomic wigeon

[tired vigil]

someone said before and it worked for me in the past "Shy" does the job pretty well

it makes them look at the camera but not directly

also "thinking" but I never tried that one too much

Some shy examples

[atomic wigeon]

If you use "commercial pose" or "athletic pose" you'll get similar expressions, but with more body pose variety.

[tired vigil]

that one I had model in the prompt so it looks like... a model lmao

[atomic wigeon]

The word "pose" automatically gives you a professional model. The reason for that is normal people don't typically "pose" for the camera. So that one word will typically force the AI to give you a professional model, in their 20's, athletic and healthy, and even select locations where people typically pose for camera's.

[ashen goblet]

need some help with a clothing style that the ai keeps getting wrong for me. I want to use fishnet clothing, but the ai does generate it.. alongside fishes and aquariums 😭 , already have all that stuff in negative prompt. Is there maybe a better positive prompt for that kind of clothing?

[tired vigil]

Have you tried a lora ?

[next flint]

so idk what model you're using and if this would work for people but i remember on the wiki they said you can say things like up 40 left 60 for the camera. maybe you could apply that to where the character is looking

[ashen goblet]

don't have any for that clothing style

[next flint]

are you using a ui that allows you to download them?

[tired vigil]

Can you show me what you want to replicate ?

[next flint]

so i just tried this and i dont think it works and i dont see it in the wiki so i think im imagining things

Kyomaku | 響真 i found some good results by just specifying fishnet clothing

[ashen goblet]

u know the chara designs from naruto? (if not, a quick google of fishnet clothing + naruto does the trick) a lot of them are wearing random fishnet clothing, for some legwear, for some tops, etc, almost as if its armor.. lol

[ashen goblet]

hm.. which model are you using? maybe my model just doesnt understand it. Well, it half does, ist just keeps generating actual fishes in addition to it..

[next flint]

yeah i got something with the same kind of overall look by specifying "lewd,nsfw"in the neg because fishnet

[next flint]

i was using darksushimix

[ashen goblet]

yes fishnet has a tendency for nsfw xD

[next flint]

yeah so to ge normal clothing go with those things in the neg

can you send an example of what you are getting

[ashen goblet]

(aquarium, fish:1.5), (ocean, sea, water:1.2), animals, wildlife, wings, plants, vegetation quite literally my negative prompt addition atm in an attempt ot make fishnet work

[next flint]

what model are you using also

[ashen goblet]

these 2

[next flint]

so i got this when using a basic prompt
1 person, standing in office, fishnet clothing
with the neg
nsfw, lewd, (aquarium, fish:1.5), (ocean, sea, water:1.2), animals, wildlife, wings, plants, vegetation

the colors are screwed up likely cuz the neg in some way

[ashen goblet]

eh, I'll DM an image, since it's a bit on the border to nsfw 😓

[next flint]

k

[ashen goblet]

still safeish tho xD

[next flint]

i would do the same. seen lots of people get it removed just cause it was close

[tired vigil]

send me too

[next flint]

do you want me to send it to you

[tired vigil]

Sure

I'm intrigued now to try the random fishnet style clothes, seeing if it works haha

[next flint]

it definitely would have been taken down in this server

[tired vigil]

I was googling it, looks like they aren't fishnets but some sort of chainmail

[next flint]

that makes so much sense

[tired vigil]

the more "ninja" version of a chainmail haha

[next flint]

but i think if you told sd chainmail it would just be armor

maybe "metal fishnet clothing?

[tired vigil]

Doubt the AI would get that

[next flint]

sdxl would

cuz its supposed to have that thing where it can come up with new ideas

[tired vigil]

hmm

I think regional prompting might do it, tiled diffusion

[next flint]

ill try with both sdxl and sd 1.5 and see what happens

[tired vigil]

idk whats that called exactly

[next flint]

yeah

[tired vigil]

cause getting random areas with fishnet clothing seems too hard haha

[next flint]

actually ill try chainmail clothing

okay i found a reddit post specifying that its kusari katabira

ill try that too

[ashen goblet]

exactly. I gen a lot of art for D&D, and any kind of attempt at generating armor usually goes into heavy plate / scalemail stuff lol

[next flint]

oh then r u using the dnd model on civit?

[ashen goblet]

nah, just random anime models

my favourite trick recently was + using ghibli/hayao miyazaki as a style prompt to generate an artstyle like this.. but I still need to improve this image, some stuff needs fixing

[next flint]

so i was able to get this with some basic prompting with sdxl

it does not like to use kusari katabira

[ashen goblet]

that's a pretty good chaiunmail

[tired vigil]

the armor lmao

[ashen goblet]

that turned out well, but yeaht he armor is random xD

[tired vigil]

is that what you want kinda of ?

like

[tired vigil]

I can't get it to not generate stockings

[ashen goblet]

basically, I guess, but yeah, would be good to get the AI to understand where you want fishnet clothing and where not

maybe the trick is img2img?

drawing a poor version of it and praying the ai understands

[tired vigil]

I think tiled diffusion would do it no problem

you can specify areas of the image to do the prompt alone there

[next flint]

okay i think its getting closer

[ashen goblet]

havent heard of that before

[next flint]

i doubt it will work but im going to try instruct pix2pix and see if i can make normal fishnet clothing turn into metal

[tired vigil]

@ashen goblet https://github.com/hako-mikan/sd-webui-regional-prompter

[next flint]

what do you mean regional could do the chainmail

okay the prompting on this damn thing was very annoying

[tired vigil]

I mean if you're going for anime style like in naruto they look more like fishnets than anything

[next flint]

honeslty im thinking just do fishnet and call it "black steel"

or something like that

cuz chainmail fishnet aint working and sd doesnt like japanese

[atomic wigeon]

Try bubble wrap. Geometric pattern. Geometric MPAP.

For something that specific I would use Davinci Resolve and make the pattern to overlay my image.

And then let the AI repair it and merge it all.

[next flint]

i think i gotta just wait for controlnet cuz i was able to get the undersuit correct but it struggles to add normal clothes on top

but ill def try those tokens

[limber scroll]

Any way to get deterministic comfyUI wildcard prompts? Same seed produces different prompts 😦

[next flint]

Any recomendations for getting a photo that looks real? ive tried things like RAW photo and a few other things indicating a photo but it still looks a bit too idk smooth

[atomic wigeon]

Add a specific camera and lens.

For example: Sigma fp with Sigma 45mm f-2.8 DG DN

[next flint]

alright. normally i just include 24mm

[atomic wigeon]

https://rikkar69.github.io/SDXL-artist-study/cameras/

young woman wearing leg stockings made of polished aluminum fishnet centered symmetrical

Is that what he was looking for?

[next flint]

he was looking for like kusari katabira

[obtuse torrent]

IDK ... he never show a example... maybe fishnet pantyhouse?

[atomic wigeon]

I dont know what that is

[next flint]

its the chainmail thing in this image

idk he used naruto as example

[atomic wigeon]

So basically see thru fishnet armor?

[obtuse torrent]

and kusaragi is the name of this japannese armor?

[next flint]

its chainmail but thing and like a shirt

[obtuse torrent]

kusaragi is a old booru tag too.. i think ..sorry was kuzasuri ..:P

[atomic wigeon]

Ok

I'm about to get fired at by an NSFW machine gun.

[next flint]

Wdym

[atomic wigeon]

NSFW's galore

Trying to make that

[prisma grail]

I'm trying to make a person squatting in sdxl 1.0, what's up with the grayscale?

I'm doing something wrong

[atomic wigeon]

You have to give it an art style. It's giving you random results because you didn't tell it to do otherwise.

[prisma grail]

ohh

where do I get art styles?

[atomic wigeon]

https://rikkar69.github.io/SDXL-artist-study/art-mediums/

[prisma grail]

thank you

[atomic wigeon]

Np

If you're looking for a photograph you simply add a camera lens (also at that site at the top).

The lens will define the style

By default.

[prisma grail]

the lens?

[atomic wigeon]

Yes. The AI recognizes camera lens and assumes you want photographs of a specific quality and aperature

[prisma grail]

like the keyword?

[atomic wigeon]

Copy and past this. Add it to your prompt. You'll see what I mean.

Sigma fp with Sigma 45mm f-2.8 DG DN

Lens can be used with surrealism, cartoons, 3d mandalas, you name it. But a lens by itself will often produce a professional real human photograph.

[prisma grail]

Do I add it to the prompt?

[atomic wigeon]

Yup

Then wide shot, closeup, medium shot, extreme long shot, birds eye view, Dutch view, etc. Those will tell the camera where to be in the shot.

[prisma grail]

alright

[atomic wigeon]

Natural light, natural shadows to seal the deal. Or motivated light.

Athletic poses typically only use motivated light or dynamic lighting.

[prisma grail]

alright

a lot to take in but it's interesting

[atomic wigeon]

https://www.youtube.com/watch?v=Ng_GmJy_F8c&t=422s

Ignore the MJ. It applies to SD too

[prisma grail]

that video is comical

[next flint]

CreativiZt for the site you provided using the film catagory should i add film after i copy the film type?

[atomic wigeon]

It's personal preference. You can use one or the other, both, or any combination. So for example, you want to hover a drone over a woman on the beach. Get a close up from above her. You can combine camera lens and views.

Prompt that uses two view to focus on my target subject:

woman in bikini laying on a towel extreme close up crowded beach birds eye view 3d Ricoh Theta Z1 with Built-in 14mm f-2.1

[next flint]

oh i always put camera stuff at the beginning

let me try x/y/z plot and see how it changes

[next flint]

or at least i tried putting it at the beginning a few hours ago and been doing it since

[atomic wigeon]

Priority can be adjusted based on how stubborn or cooperative the AI is being with your scene. Adjust to focus more emphasis.

Most of the time lens can be put at the end

[next flint]

yeah ive always heard things at the start take priority unless a token later on has a overwhelmingly higher weight so if i wanted a really realistic image i would just throw the camera specs and shot at the beginning with a weight of 1.25

[atomic wigeon]

It's mild in most cases. Especially with SDXL 1.0. But it's true.

[next flint]

the one on the left has it at the beginning and the one on the right has it at the end. otherwise same exact parameters

it doesnt seem to be a huge difference

[atomic wigeon]

Exactly. It's a mild zoom. But a seed randomizer or manual change to digits can do the same thing

[next flint]

same prompt different seed except now the one on the left has the camera stuff at the end

i think its basically negligible

[atomic wigeon]

Medium shot in prompt?

[next flint]

nah

[atomic wigeon]

That would force that specific frame

[next flint]

aight let me try adding that

[atomic wigeon]

I typically put medium shot or extreme closeup at the beginning of the prompt

extreme long shot too. That one is finiky

[next flint]

alright i added medium shot. left is at start and right is at end. mod plz dont delete right img

[atomic wigeon]

Yup that's a typical human medium shot

By definition

[next flint]

im going to try it with a different film type rq

just cause i want to make sure im not basing results on too small of sample size

[atomic wigeon]

Try motivated light. It may look better with that seed.

[next flint]

i added a bit to the negative since i was getting object clipping

okay BIG difference. on the left the stuff is at the front

im thinking ill try maybe just moving camera shot to the back for both

[atomic wigeon]

Left side is a Closeup. Right side is medium. Promptable.

[next flint]

okay in that seed something is weird. ill try a different one

i moved the shot to the back but it still made it closeup

[atomic wigeon]

Extreme closeup will zoom on the left image. Knee high shot will zoom out on the right image. Or full body shot.

Sometimes if you add "Dutch view" it elongates the shot to the next tier lens zoom.

[next flint]

okay for some reason even when changing the seed im still getting closeup on the left

although i gotta say it is actually looking like a real photo (besides the typical ai issues)

[atomic wigeon]

What lens did you use?

[next flint]

Kodak Ektachrome 64 film

this is my stuff to make it more like a real photo
(Kodak Ektachrome 64:1.25),(photo, real, real shadows,motivated light, natural shadows, real photo:1.15)

this is from the site

[atomic wigeon]

Nice. The white background is not very good with that one. Try "muted color background".

That one is better.

[next flint]

DAYUM

like its kinda too good in some areas

so its easily noticeable as ai but the lighting is rly good

[atomic wigeon]

Add:

sharp focus, intricate detail, flawless clarity, cinema quality

Those are mechanical trigger words

[next flint]

how about i just add the rest of my photo prompt
4k textures, soft cinematic light, adobe lightroom, photolab, hdr, intricate, elegant, highly detailed, sharp focus, (cinematic look:1.4), soothing tones, insane details

i got the one on the left with the things i added and the one on the right with yours

[atomic wigeon]

You can also add "soft and fluffy hair", "add static to scene" or "wild hair". They will tweak her hair a bit

Soft and fluffy hair is great for animals too

"random face pose"

[next flint]

would be nice if the webui i am using had wildcards

[atomic wigeon]

The word random is a wildcard

Add it to anything

[next flint]

rly?

cool

[atomic wigeon]

Random rustic background, random pose, random face pose, random expression, random eyeglasses style, random mood, random vibe,

Random will trigger priority at the beginning of prompt and work backwords

random time for dinner

Really. Gotta run lol. Take care bud.

One last tip. Add spotlight and color.

[next flint]

you too. thanks for the help

[atomic wigeon]

I guess we're ordering pizza. Anyway, here is that prompt. Notice no specific details on lighting or quality. SDXL 1.0 doesn't really benefit from all that anymore. Rare exceptions.

Extreme closeup face shot of a healthy young Finnish women with freckles and beauty mark Sigma fp with Sigma 45mm f-2.8 DG DN, blue spotlight

[next flint]

Really. But should i keep my natural light and natural shadows in my prompt?

i mean im assuming it cant hurt but ik lower tokens is more likely for better images

[tired vigil]

#1100170312106127410 message

so is there a way to make it generate 2 distinct characters?

i know it's difficult

i also need to isolate morty and get him right on his own i suppose

https://tenor.com/view/pink-floyd-the-wall-anybody-in-there-gif-14394079

just nod if you can hear me

is there anyone at home

[quiet zodiac]

You'd have to post your prompt, then others can help refine it.

[near glade]

Need some help in Inpainting.

I wan to remove that logo in the middle, i tried a lot of things in inpaiting but didn't work, the best i could do was to leave a "stain" there. What setting are you guys using to have it like a "magic eraser" like functionality ? Also is there a model i can use to achieve that?

[next flint]

i was able to get using instructpix2pix

[near glade]

It still is a little blurry? But better resu;ts than what i got.

https://github.com/Klace/stable-diffusion-webui-instruct-pix2pix/
It says this tool is deprecated?

[terse python]

Cant fix the many legs no matter the negative prompt

Any tips? 😫

Or like, just go on with the luck?

[near glade]

Go ham on negetive prompts OR use Controlnet and inpainting 😉

[next flint]

you dont need the extension. its just a model and the feature is now built in. i typed in "remove the logo from the shirt" into the prompt

[near glade]

I see.... let me try this out. Ty!

[next flint]

inpainting using the fill setting would likely leave a bit of a stain look but it probably looks okay

[terse python]

Thanks for the tip! I think for these I'll go for inpating, prob will be better, or maybe turn it to canny then remove it in some pait-like software?

[next flint]

this was the closest i could get

[near glade]

This is exactly what i wanted. What was your workflow?

[next flint]

id recommend trying some other options for negative considering you also have 2 watermarks

[near glade]

So found this model, that exactly does the magic eraser funtionality
https://github.com/Sanster/lama-cleaner

[frigid grove]

Getting 2 characters while using RP is pain

[next flint]

i mean there are lots of ways to inpaint it using ai. i was just trying to think of ways to do it in the webui

this was done with clipdrop using the cleanup feature

[quiet zodiac]

@terse python try lowering your resolution or setting your resolution to one of the power of 2 values.

[atomic wigeon]

Light inside of a scene gets scattered into RGB's (red, green, blue). It's the science of color. Browns are a result and often lead to noise and pixilation. So when cleaning up images in post you need to consider the artifacts and remnants of an area of your image. Extend your boundary and make smooth round brush strokes in a much larger area than what you're clearing. This is why people struggle with background removal and staining.

[atomic wigeon]

Shoes. They come in pairs.

[rapid dew]

Hey there, not sure what to do or what I'm doing wrong but the bot just refuses to generate a bulletproof vest on an FBI agent even though I tried to put emphasis on it

[hoary galleon]

So the bot is refusing to generate this "1988, a blond taned woman, holding a gun, wearing 80's winter clothing, black jacket, red gloves, covered by blood, running scared, nigth time, city ligths, atmospheric,police ligths, by russ meyer" in a realistic setting, no matter what I add to the promt, it continues to make this poster-like pictures(? when in reality I want something more realistic. Images below to understand what I'm talking about, any help?

[next flint]

does the bot have the option to do negative prompts?

[hoary galleon]

how do I type the negative prompts to the bot?

[next flint]

idk i was wondering if you could

[hoary galleon]

i'm so lost here I'm kinda new so

[next flint]

so are you looking for a real photo style?

[neon patrol]

Ross meyer made exploitation films, and you're giving it a year. Clearly it's making some assumptions (and honestly doing a pretty good job)

Try something more like

a cinematic still from a movie set in the 80s of a blonde tanned woman with a gun wearing winter clothing and black jacket, red gloves covered in blood, running scared at night being chased by the police, flashing lights, photo, realistic, RAW photo

and also check for typos 🙂

Could use a refiner pass

(that's comfy, not the bot.. but a dead simple workflow)

[hoary galleon]

thanks!

[atomic wigeon]

The word you're looking for is "style". So it would be "1980's style". But more importantly is "photograph" or "cinemaphotography".
The AI is giving you exactly what you're asking for. You have to define the medium, or it will continue giving you illustrations, pop art, 3d modeling, etc.

And for the record, that "illustration of a young beautiful 1980's style blond woman with bluejeans a black leather jacket red gloves standing in the middle of a busy street with a city background and holding an oversized Desert Eagle 50 cal at night tine with motivated light" is pretty badass. I like it!

'
Speaking of "badass", it's a cool one word prompt. Try it out.

[hoary galleon]

thanks! i finnaly got what I wanted, I think this looks "pretty badass" lol

[orchid ore]

you mean small like model? Try model of /small/ garden.

optional

vector graphics need special aproach, didnt know it

no it does not @vast jetty

it was optional just in text, not in prompt

[atomic wigeon]

You could try to focus on a tiered approach. Like your first image would be a small garden the size of a bedsheet, or on top of an object aprox. the size of what you want. And then add draught conditions.

Then with that image you import it back into image to image and say lush healthy small garden.

There is a way to get that in a single prompt, I'm sure of it. But my coffee is brewing and I'm morningdumb atm.

😂

You could try the word organicpunk (as one word) as well. It may be enough to recognize the garden and generate very little fruit/veggies.

In the spring is another idea. Spring is when planting and sprouting begin.

Maybe even a combination with "foodtography style". The word style is important. But it's a mechanical camera lock that will randomize color variants in a scene. In this case, with the right order of words, it may deliver very light vegetables to plants.

Well punk added to a word amplifies many aspects of a scene. Organicpunk is a popular AI hair alteration and other cool fun things.

[quiet zodiac]

@vast jetty You could try adding the terms isometric tilt-shift to your garden prompt. Or make it a rooftop garden or backyard garden to help define the size.

[atomic wigeon]

It must be organicpunk, not organic punk. Two different mechanics.

[atomic wigeon]

Rooftop garden is a great idea! They're always small. Well, most are. Greenroofs are getting kinda big right now. Even Ford has installed them on all of their factories.

The AI has trigger words that you wont find online. Punk, hyper, wave, etc. You add them to words to trigger a unique, AI only, type of art style.

Little tips like this, and their discovery, come from small AI circles and professional artists. In some cases they sell their art and guard the tricks.

I have no idea how the technical stuff works. I'm jsut a dumb artist.

[orchid ore]

i found flowerpunk and biopunk in clip itterogation 🙂

[atomic wigeon]

There are videos on AI prompt tips. I highly recommend them.

In some cases there are hundreds of one word prompts.

Badass is one of them

As far as Youtubers go I would recommend "Making The Photo". I don't like her, at all, but she's a very good artist and has some great tips!

Outline = sticker, paper cutout, 3d paper layers, and more. But that best out of all of them is Eiko Ojala. He'[s an amazing artist! Type "Eiko Ojala style" to see what I mean.

You're gonna drop your jaw. It's sooooo good!

Yes. Using a name like Van Gogh will draw images of him personally and/or his actual artworks. By using "style" you force a scne transition. Works with film directors too. Like Peter Jackson and Oliver Stone. It even incorporates their filmography, lens used, lighting, etc. So "style" can be far more powerful.

[toxic cape]

is there a way to have the character's structure fixed?

[atomic wigeon]

Back to "organicpunk" It will look different in every filter, size, subject matter, lighting conditions, etc.

[atomic wigeon]

Can you be more specific?

[toxic cape]

i want to have my character to be consistent

[atomic wigeon]

In the above image, if I add "green spotlight" or garden or anything pertaining to a food, veggy, fruit, etc. it will completely alter her and the scene.

[atomic wigeon]

I'm sorry. I dont understand what you're asking. But maybe "centered" and "symmetrical" will help you.

And a good negative prompt.

[orchid ore]

probably he meant lora, to have own character.

[hoary galleon]

Yeah that's what I want too, but I only use the server bot so idk if It's possible to do that

[orchid ore]

very unlikely, but realy dont know.

[dapper jacinth]

What keywords would be important for generating stock images?

I am using stock photography, Award-winning, trending on pixabay

[quiet zodiac]

You could try boring, generic, lackluster.

[atomic wigeon]

Random

[ripe bronze]

could someone walk me through how to make a prompt for ths type of picture?

[ancient wagon]

It keeps saying "IndexError: list index out of range" when I try to use the region manager

nvm, it just doesn't like the "common prompt" on

[next flint]

oh yeah common prompt needs to have a break so it would be
(common prompt)
BREAK
hair prompt
BREAK
hair prompt
BREAK

[ancient wagon]

https://github.com/hako-mikan/sd-webui-regional-prompter#2d-region-assignment

^this doesn't have BREAK in the common part

[next flint]

so that image uses a more complicated divide ratio

this is at the top of the page

[next flint]

i mean tbh you might not even need regional prompter

i mean it would likely make things easier. im going to try a few things with some prompts and see if it works

[atomic wigeon]

The struggles are real today.

[next flint]

do you know what the hairstyle is called?

[ancient wagon]

[next flint]

like the one you originally were trying to replicate

cuz im kinda able to get the hair color sepparation but idk what the hair style is actually called so even if i could get the colors right the style would likely be completely different

[umbral veldt]

Odd. Most of the images I'm generating are all coming out as like a grid of images of the thing I wanted. Even with wildly different prompts and styles. Are you guys getting this too? Is it something with SDXL?

[next flint]

cuz this was the closest ive gotten without knowing the hairstyle

[next flint]

CreativiZt do you know of a way to use a film type from the style site but have contrasting colors in the image?

[tired vigil]

street

[next flint]

are you using sdxl or 1.5 model?

local install?

i mean maybe this with a negative weight https://civitai.com/models/122359/detail-tweaker-xl

minimalism?

[terse python]

Yes, the watermarks is a problem in my custom lora acctualy, I'll fine tune it really soon!

[sharp elm]

what are the best resolutions to prompt SDv1.5 at?

[next flint]

its trained on 512x512

[tidal flare]

What command do i give in the bot channel to recreate a seed?

[frosty aurora]

How would you load a negative Lora in comfyui? Do you just type it like A1111? Or do you connect it similar to a positive one?

[tender spire]

Does anyone know if there’s a chatbot or some sort of AI tool made to refine prompts for stable diffusion? Because I don’t have the first clue about all the weird prompt engineering tags and “weights” and stuff and it has no idea what I’m talking about.

[frosty mountain]

You could always ask ChatGPT and see what he says

[eager bobcat]

ChatGPT doesn't know about SD because its training ends in 2021. And it's an it not a him. 😆

[rich crow]

probably a translation thing, a few languages interchange it/he/she depending on context

[stoic wave]

can someone give me a realistic goth women in a leather belted full skirt dress prompt i am stuck in a creative rut so add your own spin on it and i need it to look realistic

[next flint]

i found these worked for a good realistic prompt and seem to give different results

"(Goth woman),leather belted full skirt dress, moderate simplicity, dark and moody, black color palette,(Kodak Ektachrome 64 film, medium shot:1.25),sharp focus, intricate detail, flawless clarity, cinema quality, (cinematic look:1.4), soothing tones, insane details,(photo, real, real shadows, motivated light, natural shadows, real photo:1.15)"

This will give a dress with the bottom half leather
"Goth woman in leather belted full skirt dress, Dark aura, tall leather boots, Minimal makeup, bold deep-red lips, Dress with lace details, contrast of leather and lace, Dark wavy hair, pale face, Cinematic lighting, dramatic shadows, Dark color palette, black, purple, burgundy,(Kodak Ektachrome 64 film, medium shot:1.25),sharp focus, intricate detail, flawless clarity, cinema quality, (cinematic look:1.4), soothing tones, insane details,(photo, real, real shadows, motivated light, natural shadows, real photo:1.15)"

while this will give a full leather dress
"Goth woman in leather belted full skirt dress, Dark aura, tall leather boots, Minimal makeup, contrast of leather and lace, Dark wavy hair, pale face, Cinematic lighting, dramatic shadows, Dark color palette, black, purple, burgundy,(Kodak Ektachrome 64 film, medium shot:1.25),sharp focus, intricate detail, flawless clarity, cinema quality, (cinematic look:1.4), soothing tones, insane details,(photo, real, real shadows, motivated light, natural shadows, real photo:1.15)"

[next flint]

Should give moderately realistic results but if it isnt id recommend taking out tags like dark and moody, black color palette

[atomic wigeon]

Pose, noir, goth, realistic, Kodak Ektachrome 64 film, medium shot

or

photograph, chromatic, noir, goth style, motivated light, full body shot,

or

photograph, cybergoth, realistic, full body shot, skirt, glamour pose,

[atomic wigeon]

And one for fun:

3d chibi, badass, illustration, goth, street art graffiti, vector graphics, hip hop, full body shot, skirt, glamour pose, sexy, perfect teeth, smirk, headphones,

[upbeat steppe]

hi 🙂

how can I generate a final image with prompt using this draw, to put a stone grey wall in the grey color?

[atomic wigeon]

Image to image. Describe the building as if you were explaining to a builder how to build it for you. Be specific. Even tell it what stone to use. Tell it what style (modern, minimalistic, contemporary, rustic, etc). Tell it what color trim to put on the windows. Tell it what shrubs, trees and flowers you want in the scene. Your vision is the only limitation.

[quiet zodiac]

@tender spire Just start with well formed sentences. Check for spelling mistakes. Way too often I see misspelled words just thrown at the wall. Other than that, you can revert to smashing keys to submit garbage. Any type of input will make something. Weights are not that important unless you're trying to refine a concept that is almost there.

@upbeat steppe Use masking/inpainting with your posted image as an image guide. Then your prompt will only affect the area you have masked.

Try thinking of it as a coloring book.
Clean line drawing of a mermaid, coloring-in sheet style for children's coloring, no fill, just line
You could leave out no fill. Also I found with coloring book prompts that switching to Euler worked best.

Just try things, right? Remove "coloring-in sheet style for children's coloring" add "thick stroke outline"

If you're attempting to create a well-known children's author style, add their name to the end of your prompt. Some models might pick up on that.

Image guides are king. Just use that image you posted as a reference. Then prompt around it.

Here's my output using this prompt and supplying the vector reference at a source. I'm not sure what webui you're using, or what the correct term is. It's often called image to image.
Clean line drawing of a cute spotted kitten with a stroke outline
Here's my screen cap of your reference next to the output generated from it.

It just comes back to playing with the prompt.
Clean line drawing of a cute spotted mythical cartoon quadruped, drawn with a thick stroke outline
In this image, I'm still referencing 2% of that original cute kitten image. That's another nice perk from image guides. It helps you define placement and maintain a background color.

[atomic wigeon]

storybook style cartoon, Pikachu style, outline, cute round eyes, CONTENT HERE

add "muted color background" before cheetah (example)

0r "solid color background" or "neutral color background" or "happy color background".

a village?

Foreground and background results vary in every prompt.

Storybook style with an outline is probably as universal as you'll get. You will have to adjust and be more specific based on camera distance to target.

Outline = black
Sticker = white

It could be word priority in the prompt. And like I said, results vary based on the height, width, distance, etc. You can try "line art" or line art style" torward the end on the prompt perhaps.

Front of prompt will be excessive. Back of prompt will be subtle.

Also, thin lines and thick lines are a thing.

That could be hip hop or vector illustration combined with storybook style

I'm guessing