A compact conditioning-space adapter for the standard MiniMax H3 FL2VA and text-conditioned generation path. It was distilled from a cross-architecture semantic mapping developed with SenseNova U1.5 as a research teacher. SenseNova is not required at inference time.

This is an independent experimental project and is not an official MiniMax or SenseNova release.

Scope

Supported: MiniMax H3 FL2VA, text-to-video, first-frame and last-frame paths.
Not supported in v1: Ref2VA or other reference-conditioned workflows.
Do not use v1 for reference-audio singing or lip-sync. Experimental tests showed worse vocal articulation when the bridge was applied to Ref2VA conditioning.

Installation

ComfyUI Manager

Search for MiniMax H3 Semantic Bridge in ComfyUI Manager, install it, and restart ComfyUI.

Manual installation

Clone or extract this repository into:

ComfyUI/custom_nodes/MiniMax-H3-Semantic-Bridge/

Install the small Python dependency if your ComfyUI environment does not already provide it:

pip install -r requirements.txt

Restart ComfyUI after installation.

Adapter download

The adapter is distributed separately on Hugging Face:

https://huggingface.co/speach1sdef178/MiniMax-H3-Semantic-Bridge

Download:

MiniMaxH3_SemanticBridge_v1.safetensors

Place it at:

ComfyUI/models/semantic_bridge/MiniMaxH3_SemanticBridge_v1.safetensors

Then restart ComfyUI or refresh its model lists.

Included nodes

MiniMax H3 Image to Video + Semantic Bridge
MiniMax H3 Semantic Bridge
MiniMax H3 Clear Semantic Bridge Cache

The nodes appear under:

MiniMax H3/Semantic Bridge

Recommended settings

| Setting | Recommended value | Notes | | --- | --- | --- | | alpha | 0.10 | General starting point | | magnitude_match | per_token | Recommended mode | | alpha | 0.15 | Stronger influence used in published A/B examples |

The adapter transforms native H3 conditioning and blends the predicted semantic representation back into it. It does not merge weights into the H3 diffusion model.

Research, examples and limitations

The complete research narrative, metrics, datasets, scripts, controlled A/B examples, workflow, and detailed limitations are published on the Hugging Face project page.

This is an experimental adapter, not a universal enhancer. It may help some prompts, make little difference on others, or occasionally make a result worse. Representation-space similarity metrics are not equivalent to perceptual video quality.

License

Review LICENSE.md, NOTICE.txt, and UPSTREAM_LICENSES.md before redistribution or commercial use. MiniMax H3 and model-derived artifacts are subject to the upstream MiniMax H3 Community License Agreement.