This started as a fun experiment with one simple rule: "Nobody touches the DiT booty." 🏴☠️
The pirate insisted.
Project Chimera is built on a custom Turbo/Raw hybrid DiT, not a Turbo LoRA or simple weighted merge. The primary innovation is extensive modification of the TextFusion (CLIP) components, while the underlying hybrid DiT architecture itself remains unchanged.
⚠️ Requirements for w4a8
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
FAQ
Comments (9)
A good checkpoint; the human body doesn't easily deform, and the advantages of INT8 are obvious.
What exactly did you mean by saying "hybrid Turbo/Raw model"? Is it a Raw with a Turbo LoRA? (Then, at which weight?) Is it a merge of Raw and Turbo (then, what the ratios are?) This is important for correct prompting, and it directly affects creativity (more Raw = better creativity, more Turbo = same picture all the time, but faster).
Does it do NSFW or it it just the base model? Just asking because it's on civitai.red.
@ferrrett33 yes mostly it way more uncensored than base
is the INT8 convrot version? or regular int8? btw you have typo in int8 >inft8<