Please check out my Ko-Fi page if you are interested in supporting me.
LoRA commissions open! -> Fiverr
This is a LoHa LoRA! You need the extension! Why? They do a lot better with multi-concepts.
Trigger words for V3.08 (play with LoRa and tag weights!):
positive: <lora:ColoredSkinv309:0.7>(purple skin, woman)
Colors you can try are "blue skin", "green skin", "gray skin", "orange skin", "pink skin", "purple skin", "red skin", "white skin", "yellow skin", "black skin".
Experiment with adding "very dark" before the skin color tag to get darker shades of that color.
Also, I tested it for a little bit and it seems like it does still recognize colors not present in the dataset for example: "maroon skin" made a very nice red-brown color in a friend model.
Follow the steps in the Model Version notes to reproduce my images.
Sample images generated with AbyssOrangeMix3A1.
All negative embeds should be easy to find. and Prompts should be included in the sample images! Click the little icon in the bottom right to see it.
You might want to add some negatives to counter color bleeding if necessary, for example:
(purple background, purple eyes, purple clothes, purple hair)
It was trained on 200 512x512 images(20 for each color) scraped from the internet and 6 epochs using Animefull-final-pruned for the Training Model.
Danbooru tags generated by Waifu Diffusion 1.4 Tagger and fine tuned for specific details like nipples texture, see-through, and impossible shirt or form fitting.
Description
Generate using a weight of 0.4-1.0 like this: <lora:ColoredSkinv219:0.75>.
How I archived my generations:
txt2img: DPM++ SDE Karras with 20 steps, 512x512(can go up to 768 for wide or tall), Restore faces on (using CodeFormer weight 1.0 in Settings), Hires-fix on, to 1024x1024 (double original size) with your preferred upscaler and Denoise 0.4, CFG 7. Then send to img2img!
img2img: DPM++ SED Karras with 20 steps, Restore faces on (CodeFormer weight 1), rescale to 1536x1536, CFG 7, Denoise between 0.3-0.7 (based on how much you want to improve the image), SAME PROMPT! Send to extras when happy with img2img/inpaint results.
extras: Scale to x3 (final will be 4608x4608, I suggest going lower, I did x3 just for the samples here) Upscaler 1 - FatalAnime 4x, Upscaler 2 - SwinIR 4x with 0.10 visibility, GFPGAN 0.10 visibility, CodeFormer 0.10 visibility.
If anyone is curious about my version names: v{version}.{trainingEpochs}
FAQ
Comments (26)
Anyway to get shades of green or other colors in there? I always seem to get the same shade.
You can try messing with the "green skin" tag weight and the LoRA weight but unfortunately each model just does the shade it wants.
I wanna say I've seen "light purple skin" show up differently than regular "purple skin"
Some models have a very narrow understanding of green, while others will recognize terms like mint green, forest green, etc.
@some_cultist Agreed, I'm currently using Realistic Vision 5 and fighting it hard to get some unnatural details. Might have to switch to something else for supernatural prompts.
I cant seem to get pink to work, no matter the ratios, I get pink skin on maybe 1% of creations? Maybe Im just not understanding the ratios but Ive tried all sorts of variants, and models.
Can I get a prompt, model and if you are using other LoRAs with it? Also, do you have errors in console? Could be that your extension is not working properly
@Part_LoRAs Even if I just use the default LoRA prompt, change it to pink skin, or I use the pink skin example, copy and pasted, with minimal to no other content, the generation is mostly just "white."
No other LoRAs, have tried mostly with revAnimated and sunshinemix. No errors.
@jrjetu Do you have negatives? I should work just fine if you do "pink skin:1.3" and 1 LoRA weight. You probably have a negative TI or just negative tags that are fucking with it OR just giga over fit models(99% of the ones in CivitAI are giga over fit ~shrug~).
Are you sure you have the extension working correctly?
I found that the model I was using seemed to wash away the pink as it iterated through, and would end up just being a very light skin tone. Lowering the sampling steps made the skin more pink, but resulted in low quality images.
After some fighting with it, I put together these two prompts that seemed to work decently:
pink colored skin, [pink skin|pink red skin|red skin], pink face,
and
pink colored skin, [dark red skin::18], [pink skin:17], red pink face,
These use Automatic1111-specific syntax to mix "red skin" and "pink skin" prompts together. The first one alternates between "pink skin", "pink red skin", and "red skin" every step. The second one keeps "dark red skin" in the prompt until step 18, and keeps "pink skin" out of the prompt until step 17. The second one only really works with my models if you adjust the steps to be within 2 or 3 of your total steps. In this case I was using 20 steps, but tried it with other step counts as well. So 30 steps would be [dark red skin::28], [pink skin:27] instead.
Obviously, that can all be adjusted as you see fit. Those are just what worked best for me. I kept the LoRA at 1.0 for most testing, but with a smaller prompt I was still able to reliably get pink skin with the LoRA strength as low as 0.2.
Check out what i posted in the comments on the lora, and the image gallery i put in the reviews.
Love this, although it's definitely a fight to keep other details from changing to the custom color. Negatives are clutch.
Yes, unfortunately you need negatives to avoid the color bleeding over. For some reason, images with colored skin tends to also have hair with the same color... She-hulk artwork is a good example of green skin AND green hair. So to counter that, you have to prompt for colors of other stuff and also negative "green hair" for example, for better results.
@Part_LoRAs Have you thought of using different keywords that don't include full color names?
@ptdurnvcoedqegsdnw yes and it probably will work but I don't think it's necessary. I can easily change the tag for something different with a single click using my custom software but I won't cus I can't test the trained LoHa since Colab prohibited the WebUI and I don't have a GPU.
I know this is an old discussion but there is a handy guide I've found very useful here: https://civitai.com/models/18840/no-more-color-contamination-95percent-result
@vingar This is working great. I'm currently separating: purple skin, red lipstick, green dress, and it looks great. Thank you so much!
Use the lora control extension and trigger it later in the generation, like :
<lora:coloredSkin-08: 0@10, 0.5@11>
edit: i posted a more advanced guide in the comments on the lora after trying out that cutoff extension
I keep getting errors when i run the loha version using the additiona networks extension is anyone else having this issue?
You need to use the other extension by the same author
Is it supposed to be "grey skin" or "gray skin?" You use different things in images, description, and trigger words...
Yikes, looking at the dataset its written "gray skin" :(
English is not my first language even tho I'm very used to it.
I was probably low on sleep and ended up fucking the tags BUT it should still work with btw cus of how SD works
Looks like both are correct ways, but yeah. It was trained with "gray skin".
@Part_LoRAs Sounds good. Thank you!
Is there a recommended way of forcing skin color to face? I'm trying on a specific person who stable diffusion models inherently
Not really, w/e skin you prompt should already color the face. Use a higher tag and/or LoRA weight if necessary: "(red skin:1.3)".