ClipProj-MiniMax-H3

ClipProj-MiniMax-H3

Lets a small Qwen3-VL stand in for the 32B text encoder of MiniMax H3 in ComfyUI, cutting VRAM from 15.7 GB to 4.5 GB. A proof of concept that needs a custom node.

Made by NicoLab28 on Hugging Face. These notes are Mindket's, written from its model page and model card; the model's own files and license are on Hugging Face.

About

What it is: ClipProj is a pack of projection matrices for MiniMax H3 in ComfyUI. With them, a compact Qwen3-VL (4B or 8B) reads your prompt in place of the 32B Qwen3-VL that MiniMax H3 normally uses. Nothing else in your workflow changes. The author reports that VRAM falls from 15.7 GB to 4.5 GB.

Who it suits: It suits ComfyUI users who want text-to-video with MiniMax H3 on a graphics card too small for the 32B encoder. It also suits people curious about how a diffusion model uses its text encoder.

Industry | Who uses it | How they use it | Example
Video creation | ComfyUI users with limited VRAM | Swap the 32B encoder for a small Qwen3-VL plus a projection file | Generate a text-to-video clip on a card that cannot hold the 32B encoder
Localization | Creators working in several languages | Use the v3.1 files for better pronunciation in non-English speech | Make a clip with a character speaking Spanish or French
Research | Model tinkerers | Study a linear map between two encoders trained apart | Compare the 4B and 8B files against the 32B

Anyone without ComfyUI, or anyone who needs a finished tool, can skip it. The author calls it a proof of concept.

Where to use it: The files only work inside ComfyUI, through a separate custom node.

• ComfyUI, with the ComfyUI-ClipProj custom node installed
• Windows 11 with an NVIDIA card, the setup the author built and tested on
• Hugging Face, where the files are hosted

Use case | Who | What they get back
Run MiniMax H3 on less VRAM | ComfyUI users | The same workflow with a small Qwen3-VL as the text encoder
Improve non-English speech | Multilingual creators | Fewer phoneme errors than the earlier v3 files
Compare encoder sizes | Researchers | A 4B and an 8B file measured against the 32B

How to set it up: You need a working ComfyUI and the custom node before the files do anything.

• ComfyUI, with MiniMax H3 set up
• The ComfyUI-ClipProj custom node, version 0.1.13 for the v3.1 files and 0.1.4 or later for the celeb-mlp files
• A Qwen3-VL encoder, 4B or 8B
• Python, for the Hugging Face hub tools

1. Install the ComfyUI-ClipProj custom node.
2. Install the Hugging Face hub tools.
3. Download the model files.
4. Pick the file that matches your Qwen3-VL size.
5. Load it through the custom node in place of the 32B encoder.

pip install -U huggingface_hub
hf download NicoLab28/ClipProj-MiniMax-H3

The part of the card shown here is cut off, so it does not list node names or folder locations. Follow the custom node's own notes for those.

How to use it: Start with mmh3-4b-ClipProj-v3.1, which the card recommends. It adds 26 MB of projection, or 4.6 GB with its encoder. Then run your usual MiniMax H3 text-to-video workflow.

Tool | What it does
mmh3-4b-ClipProj-v3.1 | Projection for a Qwen3-VL-4B and the card's suggested first try
mmh3-4b-ClipProj-v3.1-mlp | The -mlp variant of the same projection for a 4B
mmh3-8b-ClipProj-v3.1 | Projection for a Qwen3-VL-8B
mmh3-8b-ClipProj-v3.1-mlp | The -mlp variant of the same projection for an 8B

Example: You own a 4B Qwen3-VL and a card too small for the 32B encoder. You load mmh3-4b-ClipProj-v3.1 through the ClipProj node and ask for a short clip of a character speaking Spanish. The author says swapping the encoder shifts pronunciation about as much as changing the seed does.

For newcomers and experts: If you are new to ComfyUI, get MiniMax H3 running first, then add the custom node and try the 4B file. Read the card's notes on which files are current, because older tutorials name files that have moved.

For someone who knows the field, each file is one linear map between two encoders that were never trained as a pair, fitted directly rather than by gradient training. The author normalizes every score against seed-to-seed variance, since the 32B re-pronounces 5.8 of 75 phonemes differently when only the seed changes. The four v3.1 files change between 6.4 and 7.0 of them. The author also warns that cosine similarity cannot see countable attributes.

Your text prompt
Small Qwen3-VL encodes it
ClipProj matrix maps the output
MiniMax H3 diffusion model runs
You get the video

Good to know: The v3.1 gain is in speech only, with no measurable image gain over v3. Phoneme errors fall by 29 percent overall, and by 60 to 74 percent in Spanish, French, German and Italian.

Title: VRAM for the text encoder step
Qwen3-VL-32B encoder: 15.7 GB
Small Qwen3-VL with ClipProj: 4.5 GB

The author built and tested on one setup, Windows 11 with NVIDIA and ComfyUI 0.31.0, and explored only a little. The benchmark video sounds rough on purpose, with low resolution and few sampling steps.

• License: MIT
• Cost to run: none named, but you need ComfyUI, the custom node and a GPU with enough VRAM
• Data it reaches: it downloads files from Hugging Face. The shown card names no account, web or clipboard access.
• Not for: use without the ComfyUI-ClipProj custom node, or as a finished tool, since the author calls it a proof of concept

On Hugging Face, 13 November 2026

• Task: text to video
• Library: comfyui
• Downloads in the last 30 days: 0
• Likes: 180
• License: mit
• Based on: Comfy-Org/MiniMax-H3
• Tags: comfyui, minimax-h3, text-to-video, qwen3-vl, text-encoder

Links

• Model: https://huggingface.co/NicoLab28/ClipProj-MiniMax-H3
• Files: https://huggingface.co/NicoLab28/ClipProj-MiniMax-H3/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download NicoLab28/ClipProj-MiniMax-H3

Then follow the usage in its model card. Nothing in this kit runs by itself.