Meet the child of Project Chimera and Krea 2 Center.
š§ Brain: Project Chimera (txtfusion layers)
šŖ Body: Based on Krea 2 Center's DiT backbone, with additional refinements.
Chimera provides the model's prompt understanding and interpretation, while the image generation backbone builds upon Krea 2 Center. The result is a model that inherits traits from both parents while developing its own personality.
ā ļø Requirements
Please make sure you're using ComfyUI v0.31.0 or newer. Older builds may not include the required W4A8 support.
Low-VRAM users: I also made a W4A8 quantized Qwen 3 VL 4B text encoder, only 2.86 GB. You can find it here: https://civarchive.com/models/2843533/qwen-3-vl-4b-w4a8?modelVersionId=3210230
Description
int 8 ConvRot with Adaround using Prodigy
FAQ
Comments (15)
Yay! int8 - which to use? alt or not alt?
hehe up to you both work different I cant decide ;)
@xFennec777 the Unicorn example is a good one it shows they "think" quite different
@TiwazMĀ Fair enough! I will take a look at both and see what suits my needs best!
Ok, they are both really good on the multi-colour lady. Unique. I will have some fun testing the next few days. You have too many good models @TiwazM
The model has good features, understands what is generally expected of him (the images that I uploaded to the gallery are intended to show what the model is capable of), but has problems with his fingers. And he often ignores the instructions in requests to depict a dressed woman if the text contains the words big breasts or silicone filled breasts. For some reason, the woman is half-naked. I also noticed that the model incorrectly creates anatomy if a person is needed in full growth - sometimes he tries to fit the whole body into the frame illogically. But after several attempts, he manages to agree.
I have similar problems when I use the Int8 alternative version. Sometimes I got 2 or 3 persons instead of one. In my opinion this version is much more 'random' (or creative?). Which sometimes is good.
I tried to create Asien people which is still impossible for me.
@PressydentĀ I think, model still need some adjustments. Instead of doubling of persones, model time to time generate peoples with 6 fingers. Hand or feet. With Asean people i dont have problems. I would generate some more images with Asean peoples and share.
@herkus_baronas631Ā I mean, yes, there is an asien touch but ... not much.
@PressydentĀ Shared some images only with Asian people. Its what you talking about, or you imagine something else?
@herkus_baronas631Ā The portrait Geisha, chinese and japanese was what I had in mind. The other geisha is what I meant. Sometimes it looks not that much chinese or japanese. Would you share your workflow or at least the prompts?
I tried two quick portraits myself with simple 'asian'. Maybe I have to be much more specific with more context.
@PressydentĀ @PressydentĀ I always use simply prompts. No chatGPT or something like that. Most my prompts - wroten myself. Something like that
"Japanese beauty, dressed and adorned in traditional Japanese clothing and jewelry. Slim, extremely beautiful. The beauty poses in a garden of blooming sakura. Mount Fuji is visible in the distance.
@PressydentĀ If you want, i can generate some more images and show prompts. I just simply lazy to wrote prompts into picture desciption. But if it helps, ok, i would paste some prompts.
@herkus_baronas631Ā Ah. Ok. No. It's fine. I just wanted to know what my mistake could be. And if it could be the prompting. Thank you.
About the model - conclusions.
Although the model is generally not bad, some words mean nothing to him. Part of the prompt "She wears a backless crimson dress that exposes her lower back and the curve of her waistline, while the front is covered by a sheer fabric overlay. This pose highlights how the silicone implants create a smooth, rounded shape from behind. " The words "silicone implants" mean nothing to the model. The shape is natural, without any hint of artificiality. The same applies to facial features - no 'improvements' are known to this model. The only 'improvement' that the model understands is "bimbo lips". But then most often a woman with bimbo lips is wide-mouthed, unless it is specifically indicated that she should smile or have her lips pursed. It was impossible to 'dress' the woman in a completely innocent description - "an Indian beauty is getting ready for a wedding, dressed in a traditional dress, without makeup or jewelry." The model was always depicted bare-chested.
In short, the model has potential, in many situations it is very good, but in some situations it does not listen at all.