Qwen for MiniMax H3 (QP)
This model has been quantized from a 100GB source QP training.
Only 2.2GB larger for NVFP4
This model HAS NOT been ablierated or uncensored by traditional means and is unlike any other model available currently
The attached .json is a workflow that speed my IT time up by 60%
NOTE UPDATE YOUR COMFY & CUDA
Consider using --disable-dynamic-vram
pytorch version: 2.13.0+cu132
xformers version: 0.0.35
ComfyUI version: 0.35.0
comfy-aimdo version: 0.5.5
comfy-kitchen version: 0.2.34
Description
FAQ
Comments (8)
What's QP?
How is it different from the normal encoder?
Why would one use it?
So it’s not uncensored but nsfw?
Interesting concept but using the linked workflow and suggested models CTD's my comfyui build. Obviously a million variables there but never had another H3 workflow do that. Replacing the encoder with the stock official NVFP 4 version gets me father along workflow before it produces the same result
这个在nsfw方面会更有出色表现么
It is not compatible with my ComfyUI/NVFP4 setup. My Comfy is 0.33.0-13-gb963f4ad, and PyTorch is 2.7.0+cu128.
What's the benefit of this text encoder?
Why not release it in int8, w4a8, or bf16?
I have a weird issue. I can use it just fine for text to vid, but when I try to use a reference image the whole thing crashes.
Used your workflow and standard models and it produced a high quality result. However, prompt adherence seemed to be close to zero. Stuff just happened. Is that the intent? This was from a single starting frame, not a first/last, bypassed the second image and second image resize node.