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
Patch notes
Improved clothing behavior.
Donāt use caveman prompts and expect tailoring.
āHugeā means HUGE.
Some edge cases remain because apparently fabric has limits too. š”š¤·
FAQ
Comments (10)
v1.5 - FANTASTIC! (Int8 incoming?)
- Respects my character LoRA with only MINOR face and body differences (though, it's either #1 or #2 for character respect in the 20+ checkpoints I have tests - top marks!).
- Does wonderfully with my style LoRA
This is really close to a Center v3 and Chimera v1.2
My own personal complaint: mxfp8 runs twice as slow as int8 on my wimpy rig.
@xFennec777 yes int8 for sure. w4a8 not sure, need to test something. Vram wise maybe w4a4 makes more sense in that regard.
As for MXFP8, it tends to make the most sense on newer NVIDIA hardware, especially Blackwell/50-series, but Comfyās MXFP8 kernel path still isnāt as optimized as NVFP4 right now, so slower performance can be normal
Happy to test the w4a8 and w4a4 - Your models are the top ones for me. Don't mind being a low VRAM tester. @TiwazM
Chimera, Center, and ChimeraXCenter are the mains right now :D
@xFennec777Ā hehe ok, maybe I will make both. From a quick test and only Vram check it seems w4a4 and w4a8 are pretty close maybe 0.5-1 GB difference but only had a look at windows task manager not super exact numbers
@TiwazMĀ Makes some sense - it offloads some calculation + the model data.
@xFennec777Ā added the int8 with adaround
more later but making an fp8 that takes some time ;)
@TiwazMĀ All good, I appreciate you!
V1.5 looks SOOOO promising. I really hope OP can release INT8 version for 30/40 cards. Thank you in advance.
@925_Studio yes its uploaded
@TiwazMĀ Much appreciate.