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
some more messing with the clip layers only, found some things v1 did not do as expected v1.2
FAQ
Comments (7)
Were there any other LoRAs merged especially bypass or NSFW. I am asking so that I don't double it up with those LoRAs when generating
not 100 % sure as I used part of the clip from the center series, but either way its just the clip that is modified not the DiT (main part of the model) so it should be fine. The main layers is basically a mix of the original raw and turbo.
Is it possible to download fp16/bf16 of your wonderful model somewhere? Unfortunately I can't use quantized versions on my setup yet.