Hugging Face's logo

Abiray
/
Minimax-H3-nvfp4-INT4-INT8-Convrot

Image-Text-to-Video
Diffusers
Thai
text-to-video
image-to-video
video-to-video
text-to-audio-video
multimodal
quantized
comfyui
int4
nvfp4
Model card Files
xet
 Community
19

smaller text encoder?

#3
by jimzlf - opened
Discussion
jimzlf

is that possible to have TE in a smaller quantization like int4😂 it's even bigger than the diffusion model
Abiray
 Owner

Yeah, give me some time
jimzlf
•
edited 22 days ago

Yeah, give me some time

oh, sorry but I saw one here😂nearly as big as nvfp4 https://huggingface.co/Merserk/MiniMax-H3-INT4-ConvRot/blob/main/qwen3vl_32b_minimax_h3_int4_convrot.safetensors
wsq194

I tested a Qwen3-VL-8B encoder with MiniMax H3 using an experimental projection adapter, and it works locally.

Setup:

  1. Put qwen3vl_8b_nvfp4.safetensors in ComfyUI/models/text_encoders/.
  2. Put mmh3-8b-ClipProj-celeb-mlp.safetensors in ComfyUI/models/clip_projections/.
  3. Load qwen3vl_8b_nvfp4.safetensors with CLIPLoader and set type to boogu.
  4. Connect the CLIPLoader output to ClipProjApply, and select mmh3-8b-ClipProj-celeb-mlp.safetensors.
  5. Connect the ClipProjApply output to the H3 prompt/reference-image conditioning path. The rest of the H3 workflow can remain unchanged.

The 8B encoder should not be connected directly to H3. Qwen3-VL-8B produces 4096-dimensional features, while H3 expects 5120-dimensional conditioning. ClipProjApply performs the required 4096 -> 5120 projection.

This worked in my local testing, but it is a community/experimental adapter rather than official native H3 support for Qwen3-VL-8B. I would still compare identity preservation, fine details, and prompt understanding against the original encoder.

Sign up or log in to comment