README.md

---
license: other
license_name: minimax-h3-community-license-agreement
license_link: https://huggingface.co/MiniMaxAI/MiniMax-H3/blob/main/LICENSE
pipeline_tag: image-text-to-video
base_model:
  - MiniMaxAI/MiniMax-H3
  - Comfy-Org/MiniMax-H3
tags:
  - minimax-h3
  - ref2va
  - comfyui
  - diffusion-single-file
  - int8-convrot
  - synchronized-audio-video
  - experimental
language:
  - en
  - zh
inference: false
---

# MiniMax H3 Ref2VA Patch-In HF 1.02

Experimental ComfyUI single-file checkpoint derived from
[MiniMaxAI/MiniMax-H3](https://huggingface.co/MiniMaxAI/MiniMax-H3) Ref2VA and the
[Comfy-Org/MiniMax-H3](https://huggingface.co/Comfy-Org/MiniMax-H3) INT8 ConvRot repack.

> [!CAUTION]
> This is an experimental weight modification, not an official MiniMax release and not a proven
> “de-oil”, “de-wax”, restoration, or quality-fix model. The original checkpoint remains the
> recommended default.

> [!IMPORTANT]
> MiniMax H3 is governed by the
> [MiniMax H3 Community License Agreement](https://huggingface.co/MiniMaxAI/MiniMax-H3/blob/main/LICENSE),
> which defines excluded territories and mandatory redistribution conditions. Before publishing or
> redistributing this derivative, include the official `LICENSE`, keep the modification notice and
> `NOTICE`, and confirm that the intended distribution method and audience are authorized. A public
> Hugging Face repository may be reachable from excluded territories; a repository gate alone is not
> necessarily geographic access control. This model card is not legal advice.

## 中文说明

这是 MiniMax H3 Ref2VA 的实验性 ComfyUI 单文件衍生模型。它没有训练、微调或蒸馏，只对
视频 patch 输入投影中的 2×2 空间高频分量增加 2%。两组固定条件测试都出现了很弱的皮肤
高频代理正增益，但肉眼仍未确认能够消除 Ref2VA 的油感或蜡感，所以只能作为 EXP 对照模型。

发布前必须阅读上游许可证。该许可证对适用地域、公开分发、商业使用、安全措施、修改声明、
`LICENSE` 和 `NOTICE` 都有要求。

## Model description

| Property | Value |
|---|---|
| Task family | MiniMax H3 Base Ref2VA / reference-to-audio-video |
| Format | ComfyUI single-file diffusion model |
| Precision | Native INT8 ConvRot container inherited from the source repack |
| Training | None |
| Modified tensor | `video_patch_proj.weight` only |
| Input patch HF gain | `1.02` |
| Output-head gain | `1.00` — unchanged |
| File size | `34,038,894,550` bytes, approximately 31.70 GiB |
| Status | Experimental A/B candidate |

### What was changed

MiniMax H3 video latents are patchified with a `1 × 2 × 2` patch. For every latent channel, the four
spatial input columns of `video_patch_proj.weight` were transformed in an orthonormal 2×2 Haar basis:

- the DC/common component is preserved at gain `1.00`;
- the three non-DC H/V/D components are multiplied by `1.02`;
- the result is transformed back and written into the input projection;
- no output head, audio tensor, shared Transformer block, VAE, text encoder, or ComfyUI node was changed.

The equivalent 4×4 transform has diagonal `1.015` and off-diagonal `-0.005`. Its all-ones/DC
eigenvector has gain `1.00`; the three orthogonal spatial-detail eigenvectors have gain `1.02`.

This is not an output sharpening filter. It changes the video latent input projection used during
each joint audio-video denoising forward pass. It cannot reconstruct real texture that the model does
not generate.

## Checkpoint integrity

### Source

```text
File: minimax_h3_ref2va_int8_convrot.safetensors
Repository: Comfy-Org/MiniMax-H3
Size: 34,038,894,550 bytes
SHA-256: 9eef934046a0671bc8a5daf87100705e1478419c574cfde70c50fbe6885f76a9
```

The source size and SHA-256 match the LFS object published by
[Comfy-Org/MiniMax-H3](https://huggingface.co/Comfy-Org/MiniMax-H3/blob/main/diffusion_models/minimax_h3_ref2va_int8_convrot.safetensors).

### Modified checkpoint

```text
File: minimax_h3_ref2va_patchin_hf102.safetensors
Size: 34,038,894,550 bytes
SHA-256: 5552621452ff73c9627ece6bd08f08e8c08b6504ee8a753b1aa1e1abcb70d9e4
```

Structural verification:

- safetensors header identical to the source: yes;
- tensor keys and shapes identical to the source: yes;
- changed tensor: `video_patch_proj.weight` only;
- changed byte interval: `[17,879,061,191, 17,881,125,575)`;
- target writeback maximum absolute error: `0`;
- DC-sum maximum absolute error: `9.5367431640625e-07`;
- SHA-256 of all unchanged file regions:
  `82d4866debf06c9b9b7a67d645d1dcac57bdd91eba3b341a55a4e1fe9f687852`.

The machine-readable surgery record is available at [`evidence/surgery.json`](evidence/surgery.json).

## ComfyUI usage

### Required model components

This repository contains only the modified Ref2VA diffusion model. Obtain the remaining H3 components
from the official or ComfyUI repack repositories under their applicable licenses:

- MiniMax H3-compatible Qwen3-VL text encoder;
- `minimax_h3_video_vae_fp16.safetensors`;
- `minimax_h3_audio_vae_fp32.safetensors`;
- a current ComfyUI build with MiniMax H3 and native INT8 ConvRot support.

Suggested layout:

```text
ComfyUI/
└── models/
    ├── diffusion_models/
    │   └── minimax_h3_ref2va_patchin_hf102.safetensors
    ├── text_encoders/
    │   └── qwen3vl_32b_minimax_h3_*.safetensors
    └── vae/
        ├── minimax_h3_video_vae_fp16.safetensors
        └── minimax_h3_audio_vae_fp32.safetensors
```

Load the checkpoint with ComfyUI's **Load Diffusion Model** node and use a Ref2VA/R2V workflow. The
official ComfyUI template is available at
[video_minimax_h3_r2v.json](https://github.com/Comfy-Org/workflow_templates/blob/main/templates/video_minimax_h3_r2v.json).

### Recommended comparison protocol

For a meaningful A/B, change only the diffusion model:

1. generate once with `minimax_h3_ref2va_int8_convrot.safetensors`;
2. generate again with `minimax_h3_ref2va_patchin_hf102.safetensors`;
3. keep seed, reference media, prompt, dimensions, frame count, sampler, scheduler, shifts, steps, VAEs,
   text encoder and conditioning identical;
4. do not add a Turbo/four-step LoRA when reproducing the validation below;
5. inspect skin, hair, fabric, temporal consistency, identity, composition and audio together.

The local validation used the isolated
[`minimax-h3-audio-T8`](https://github.com/T8mars/comfyui-minimax-h3-audio-T8) ComfyUI nodes with
`dual_clock_euler + native_flow`, video shift `12`, audio shift `3`, and 20 joint denoising steps. That
custom node package is the validation environment, not an architectural requirement of the checkpoint.

## Validation

### Controlled setup

```text
ComfyUI: 0.31.0, commit cbbc9dab1f03d0d9a6caa8a8be7d77a7e37e1e44
Checkpoint family: full Ref2VA INT8 ConvRot
Turbo LoRA: none
Resolution: 512 × 512
Frames: 22 at 24 FPS
Steps: 20
Sampler/scheduler: dual_clock_euler / native_flow
Seeds: 2608103502 and 2608103503
```

The 22-frame test is a short diagnostic, below the official 4–15 second usage range. It does not prove
quality at the recommended 124+ frame range. One centered portrait reference and one prompt were used;
the reference image is not redistributed.

The original checkpoint was generated twice under seed `2608103502`. Decoded video was pixel-identical
and decoded PCM audio was sample-identical, confirming deterministic repeatability for this setup.

### Candidate-to-source ratios

| Seed | Skin gradient | Skin Laplacian | Skin high-pass | Positive skin ROIs | Temporal change | Audio correlation | Audio RMS |
|---:|---:|---:|---:|---:|---:|---:|---:|
| `2608103502` | `1.0325` | `1.0760` | `1.0189` | `2/3` | `1.0321` | `0.9930` | `0.9189` |
| `2608103503` | `1.0266` | `1.0418` | `1.0131` | `3/3` | `1.0134` | `0.9687` | `1.0154` |

The skin high-pass proxy increased by approximately 1.89% and 1.31% across the two seeds, without a
fixed 2×2 parity artifact. This is a weak proxy-level signal only. Manual inspection did not establish
a convincing reduction of oily or waxy skin appearance.

The first seed also showed an approximately 8.1% audio RMS reduction. H3 predicts audio and video in a
joint denoising trajectory, so changing a video-only input projection can still indirectly change the
generated audio.

Full machine-readable results:

- [`evidence/metrics_seed_2608103502.json`](evidence/metrics_seed_2608103502.json)
- [`evidence/metrics_seed_2608103503.json`](evidence/metrics_seed_2608103503.json)

### Visual comparisons

Each comparison is **source / input HF 1.01 / input HF 1.02** from left to right. The 1.01 candidate is
shown only as an additional control and is not included in this repository.

Seed `2608103502`:

![Seed 2608103502 comparison](assets/comparison_seed1_mid.png)

Seed `2608103503`:

![Seed 2608103503 comparison](assets/comparison_seed2_mid.png)

## Limitations

- This is not an official MiniMax model.
- It is not a trained fine-tune, LoRA, distillation, restoration model or calibrated de-wax model.
- Only one tensor was heuristically transformed; the direction was not learned from paired data.
- Evaluation covers two seeds, one portrait reference, one prompt and a short 22-frame diagnostic.
- No official-range 124–362 frame matrix, 768p/2K test, multi-subject test, blind preference study or
  broad prompt evaluation has been completed.
- Small proxy-metric gains may not correspond to perceptual improvement.
- Identity, framing, motion and generated audio may change even when the modified tensor is video-specific.
- Compatibility with Turbo LoRAs, third-party samplers, alternate quantization loaders and older ComfyUI
  versions is not established.
- The source checkpoint remains the recommended default for production or reproducible comparisons.

## Intended use

Intended only for controlled research and A/B evaluation of a small, auditable MiniMax H3 video-input
projection change, subject to the upstream license and acceptable-use policy.

Not intended for claims of guaranteed quality improvement, identity preservation, medical/forensic use,
deceptive impersonation, bypassing safety systems, or any use prohibited by the MiniMax H3 license.

## License and attribution

MiniMax H3 and this Model Derivative are governed by the
[MiniMax H3 Community License Agreement](https://huggingface.co/MiniMaxAI/MiniMax-H3/blob/main/LICENSE).
The upstream license, including territorial restrictions, redistribution requirements, commercial terms
and acceptable-use policy, takes precedence over this model card.

Before distributing the checkpoint:

1. include an unmodified copy of the official MiniMax H3 `LICENSE` in the repository;
2. include the provided [`NOTICE`](NOTICE);
3. retain a prominent statement that `minimax_h3_ref2va_patchin_hf102.safetensors` is modified;
4. comply with the applicable-territory and downstream-user requirements;
5. verify any commercial-use authorization requirements directly with MiniMax.

Upstream projects:

- [MiniMaxAI/MiniMax-H3](https://huggingface.co/MiniMaxAI/MiniMax-H3)
- [Comfy-Org/MiniMax-H3](https://huggingface.co/Comfy-Org/MiniMax-H3)
- [Official MiniMax H3 GitHub repository](https://github.com/MiniMax-AI/MiniMax-H3)

## Modification notice

`minimax_h3_ref2va_patchin_hf102.safetensors` is a modified file derived from the MiniMax H3 Ref2VA
checkpoint. It was created without training by applying a 1.02 non-DC 2×2 Haar gain to
`video_patch_proj.weight`. This modification is independent and is not endorsed by MiniMax or Comfy-Org.