ClipProj-MiniMax-H3
Text encoder replacement for MiniMax H3
2 min readupdated
This ComfyUI tool allows replacing the heavy Qwen3-VL-32B text encoder of the MiniMax H3 model with more compact Qwen3-VL-4B or 8B versions using special projection matrices. As a result, VRAM (video memory) requirements drop from 15.7 GB to 4.5 GB while maintaining diffusion model quality. The v3.1 update brings improved multilingual speech and support for new languages without losing generation quality.
Key features
- VRAM savings: reduces video memory consumption from 15.7 GB to 4.5 GB.
- Improved speech: version v3.1 reduces phoneme error rates by 29 percent across ten languages.
- Compatibility: runs via a custom node for ComfyUI without modifying the base model.
- Accuracy: matrices are calibrated using ridge regression with added MLP residual networks.
What you can do
- Generate video: run MiniMax H3 locally on lower-memory graphics cards.
- Synthesize speech: voice-over videos in different languages with correct pronunciation.
- Use multimodality: apply the compact Qwen3-VL for prompting and frame control in ComfyUI.
Example
Projection files must be placed in the ComfyUI/models/clip_projections/ folder and used in workflows with the updated node.
Read more
DetailsBenchmarks: cosine similarity reaches 0.8037 for the 8B version with a residual network.
Requirements: CUDA-compatible graphics card and ComfyUI (node version 0.1.13 or newer required).
Files: options available for 4B and 8B encoders (including celebrity name -celeb and MLP versions).














Comments