Edit for 4.0
More realistic, less bright and saturated though. Still very creative. Not necessarily better than the previous versions, but it offers something different.
Edit for V3.5
After several rounds of internal merges, tuning, clip editing, many almost V3 versions, I've ended up further than I thought with what I call version 3.5. It's a general upgrade that fixes many of the issues of the previous V2 and V1 model, and so far I'm not sure what are its flaws. It's real fun to play with, the model is very creative, you often have to get out of its way and let it do its thing, although it will follow prompts closely.
Quite good at low inference steps.
Minimal prompting, minimal lora use seem to work best.
Please share your creations! I love to see them.
Edit for V2:
V2 is an overall, upgrade. It's more balanced, bolder, more detailed. Subjectively, it's better than V1, but you might think differently and that's legitimate.
Original post (V1):
This is an NSFW-focused model based on ten rounds of merges. I'd love to give credits, but at this point I've forgotten the contributors, sorry.
It'll do SFW, but it'll occasionally generate sexually explicit material on its own, even without a prompt.
The goal with this model was creating an Illustrious variant which would make great semi-realistic demi-humans, without a complex prompt, without a complex Lora stack, just the model, and you're good to go.
I'm personally pleased with the results, but more than anything, I'm curious about what people think, and how they'll merge it to make something new. If you do use this model for merges, let me know so I can see what you're up to. Just out of curiosity.
All the examples images I selected were created with relatively simple prompts, no loras, and in a few second per image. No complex workflows here. I'm 100% sure people will know how to have better results than me with it.
One last thing, the model is aimed for and biased to generating mature/adult characters. I cannot advise you strongly enough to not generate minor characters. Don't do anything illegal or amoral with it.
Description
Version 9 is an attempt to rebalance the model while keeping what made it "Chaos" in the first place, a wicked creativity allowing simple prompts to come out great. I've had a lot of fun testing it, it follows prompts more accurately than before.
FAQ
Comments (21)
V9! What a return to form!
Glad you like it. I took your feedback in consideration and even tried your tongue wrapping concept.
damn it, this is probably one of the best, even at a low resolution (without upscale) the picture works out all the details, and understands complex light, works well with two or more LoRa, I tried a lot of checkpoints, but this one, I'm impressed.
Thanks! Relatively low res and low steps accuracy is important to me. I don't want to wait too long to get images when I ideate.
Actually almost any Pony/Illustrious/SDXL checkpoint can do crazy things. Key - is a corresponding prompt. Not any model know lot about anthropomorphic creatures, but... they know well about mammal/reptile/fish/insect. And about their appearance. So if you promt something like "(insect-like female:1.2) with (spider-like insect-skin:1.4)" - you can get something similar to anthropomorphic insect creature.
thankyou for a fantastic model great work, others can learn from you , i know how much hard work is involved, thanks,
Thanks for the kind words!
Can you generate some pictures that are biased towards Asian face models?
I'm not sure what you're asking me, so I'll answer some questions I do understand in the hope one will be right lmao.
Can the model generate asian faces? Yes.
Is the model biased towards generating asian faces? No.
Can I generate examples images with asian faces and upload them? Sure?
Can I make a model biased towards asian faces? Yes, but I won't, because that's not what Chaos is about.
If none of these are right... elaborate on your question :)
@_Kav_ 我尝试用一些提示词来生成亚洲人脸，但是仍是欧美人脸，请问有什么提示词可以帮助我做到这一点么
@ppzmq1234 你可以使用：(asian woman, asian girl, asian female:1.1)。举个例子。我在得到想要的结果时没有遇到任何问题。你也可以描述“亚洲人”的特征，比如直发、黑色头发等。
@_Kav_ 我试一下，十分感谢，非常喜欢这个模型的光影与质感
Hello! I have browsed all the versions of your works. I found that the facial lighting of the characters in your work was insufficient. I couldn't control the facial lighting well when using it. The characters' faces always have too many shadows.
Hey there. Chaos aims for semi-realism and aesthetic quality (subjective) on the graphical side of things, so it can't do everything well... no single SDXL model can, sadly... but that's also what makes them interesting, I suppose.
Hello. I'm sorry for replying to your message just now. Okay, I understand.
Excuse me, how can I better experience the content of your latest V10 version? It seems that your text description has not been updated. I have only been in touch with ComfyUI for a short time.
@wwwmanmoon103 It's difficult to answer. What does better mean? If it's less shadows on faces, going back to your original post, then you can alter that with negative prompting, lighting loras, and flat shading loras, or style loras.
As for the description, the main post's description bugs whenever I want to edit it for some reason, so I just update the lower part now.
thank you so much for your patient reply! 🙏
I'm still pretty new to ComfyUI and still figuring a lot of things out, so I'd like to ask you a few more specific questions:
Regarding facial lighting: You mentioned using lighting Loras, flat shadow Loras, and style Loras. Do you have any specific recommendations for ones that work well with your model? I've been searching on Civitai and HuggingFace, and there are so many options that I'm not sure where to start.
Regarding negative prompts: I've been using a fairly generic set of negative prompts before. For the issue of "too many facial shadows," are there any more precise negative terms I could add, like "dark shadows on face" or other effective combinations?
Regarding V10: You mentioned having issues updating the main post description. When using V10, are there any key settings I should be aware of, like recommended samplers, step counts, CFG Scale, or pitfalls to avoid?
Thanks again for sharing your work and for your patience. The quality of your model is really impressive, and I love it! 😊
@wwwmanmoon103 No worries.
"Regarding facial lighting: Do you have any specific recommendations for ones that work well with your model?"
It's complicated again because it's so subjective. Any SDXL type Lora should work with the model, so mostly Illustrious and Pony models, since they are derived from the same base model (SDXL). Illustrious are you best bet. Search for examples that look like what you want and try them out.
"Regarding negative prompts: are there any more precise negative terms I could add, like "dark shadows on face" or other effective combinations?"
Some words will work better than others. Order and repetition can also influence result, as well as weighting (putting things in parenthesis finished with:1.2) where the higher you go, the more "weight" you give to that part.
You usually start small like: contrast, contrasted, intense light, etc, and you add words or variations if it's still not right. It's trial and error, and you get a feel for what works and what doesn't. If you don't want something, you can also prompt for the positive. Like diffuse lighting, or flat lighting
Regarding V10: Are there any key settings I should be aware of, like recommended samplers, step counts, CFG Scale, or pitfalls to avoid?
I generated the examples at 1300 x 1300 with 15 steps, 3.0 CFG, deis sampler and beta scheduler in ComfyUI, using Euler and normal should work well in general.
Most of my generations I do at around 1kx1k, since that runs locally on my machine, that's about 6.5 seconds, and I find I can take in the image well enough and ideate quickly that way.
Going higher in steps generally improves quality and of course takes longer. I wouldn't lower than 12 steps to maintain coherent generations.
In general, I think it's good to just try things, try different models, try stuff to get an idea of what you like. Also, looking at the images you like from other people and the details of how they were generated (when that info is shared) is going to help you figure out what works and what doesn't, give you another perspective, lora ideas, models etc.
Of course, generating locally is move convenient than online, after an initial cost (Buying a GPU for AI or AI compatible isn't cheap).
@_Kav_ Thank you for your patient answers. How did you achieve a 6.5 - second image - generation speed? My computer configuration is GTX4060TI with 16GB and 32GB of RAM. It takes 50 - 60 seconds to generate an image of 1204 * 1024 pixels, including the time to load the model.
@wwwmanmoon103 I have a similar config and a 4070, but it's mostly because I don't count loading the model. Comfy keeps it loaded so it's just the first gen that takes longer. My workflow is in the images I posted, you can drag and drop the images into Comfy (if you have it installed) and It'll show you how the image was generated. Note that the showcase images for ChaosV10 was higher in Res so those take about 11 seconds to generate.