corechan/MiniMax-H3-LightVAE

MiniMax H3 — light video VAE decoder (diffusers + ONNX)

A converted copy of the light video VAE decoder of LynnReal-Omni (LynnReal-AI), made for the MiniMax H3 WebUI on Google Colab. It decodes ordinary MiniMax H3 latents with 26 transformer blocks instead of 36. Unofficial — not affiliated with, endorsed by, or supported by MiniMax or LynnReal-AI.

⚠️ Modified files. Source: stdstu123/LynnReal-Onmi-beta-0.1, comfyui/models/vae/lynnreal_omni_light_vae_fp16.safetensors (github.com/LynnReal-AI/LynnReal-Omni). We did not retrain it. We (1) converted the ComfyUI key layout to the diffusers layout (the fused to_qkv projection was split back into to_q / to_k / to_v, head-interleaved, head dim 64), (2) stored it in fp16, and (3) exported one decoding tile to ONNX (RMSNorm rewritten with basic ops for opset 17).

Files

File	What	Size
lynnreal_light_vae_decoder_fp16.safetensors	decoder.* and post_quant_conv.* with diffusers AutoencoderKLMiniMaxH3 key names, fp16 (26 decoder blocks)	3.3 GB
lynnreal_light_vae_decoder.onnx + lynnreal_light_vae_decoder.onnx.data	One tile, decoder(post_quant_conv(z)): latent latent_tile 1×24×7×16×16 → pixels pixel_tile 1×3×28×256×256; no normalization in or out (same interface as lihaoyun6/MiniMax-H3-VAE-ONNX); opset 17, fp16, external data in one file	3.3 GB
.complete	File sizes and SHA-256	—

The encoder, latent space and temporal structure are those of the official MiniMax H3 VAE; only the decoder differs. Take the encoder (and the audio VAE) from the official repository.

Usage

diffusers — load the official AutoencoderKLMiniMaxH3, keep the first 26 entries of vae.decoder.transformer_blocks, then load this file:

import torch.nn as nn
from safetensors.torch import load_file
vae.decoder.transformer_blocks = nn.ModuleList(list(vae.decoder.transformer_blocks)[:26])
missing, unexpected = vae.load_state_dict(load_file("lynnreal_light_vae_decoder_fp16.safetensors"), strict=False)
# missing should list only encoder.* / quant_conv.* weights

TensorRT — compile the ONNX file with a fixed input profile 1×24×7×16×16 (FP16), then tile the latent in time (5 latent frames + 2 overlap) and space (256 px tiles, ≥64 px overlap) as diffusers' _decode / _decode_clip do.

Measured on RTX PRO 6000 (G4), 1280×704 / 124 frames: 5.2 s with TensorRT (official decoder with TensorRT: 7.1 s; this decoder in PyTorch: 8.3 s). TensorRT output matches the PyTorch output at 52.8 dB.

License

This is a Model Derivative of MiniMax H3 (LynnReal-Omni's decoder is a distilled derivative of the MiniMax H3 video VAE decoder) and is distributed, as LynnReal-AI distributes it, under the MiniMax H3 Community License Agreement (LICENSE, a copy of the original). By downloading or using it you agree to its terms and its Acceptable Use Policy. No additional or different terms are imposed.

MiniMax H3 is licensed under the MiniMax H3 Community License Agreement, Copyright © 2026 MiniMax. All Rights Reserved.

Main points of the Community License (read LICENSE for the binding text):

1. Territory — no rights are granted to use, reproduce, modify, distribute or display the model or its outputs in the European Union, the United Kingdom, South Korea or the United States.
2. Commercial use — if your commercial products and services generate more than USD 20 million in yearly revenue, you need separate prior written authorization from MiniMax (api@minimax.io). A commercial product or service that uses MiniMax H3 must prominently display "MiniMax H3" in its user interface.
3. Outputs — do not use the model or its outputs to improve any AI model other than MiniMax H3 and its derivatives.
4. Acceptable Use Policy — Exhibit A of the license applies to every use.

---

日本語概要

LynnReal-AI の LynnReal-Omni が配布している軽量の映像 VAE デコーダーを、MiniMax H3 WebUI on Google Colab 用に 変換したものです。MiniMax H3 の潜在を、transformer 36 ブロックではなく 26 ブロックでデコードします。 非公式で、MiniMax 社・LynnReal-AI とは無関係です。

⚠️ 改変したファイルです。 出典は stdstu123/LynnReal-Onmi-beta-0.1 の comfyui/models/vae/lynnreal_omni_light_vae_fp16.safetensors です。再学習はしていません。 名前の並びを diffusers 形式に変換（結合された to_qkv を to_q / to_k / to_v に分割）・fp16 で保存・ 1 タイルぶんを ONNX に書き出し（RMSNorm を基本演算に書き換え）の改変をしています。

ファイル	中身
lynnreal_light_vae_decoder_fp16.safetensors	diffusers の名前の decoder.* と post_quant_conv.*（fp16・26 ブロック）
lynnreal_light_vae_decoder.onnx（＋ .onnx.data）	1 タイル（潜在 1×24×7×16×16 → 画素 1×3×28×256×256）。TensorRT 用

エンコーダー・潜在は公式の MiniMax H3 VAE と同じです（エンコーダーと音声 VAE は公式から取ります）。

ライセンス: MiniMax H3 の派生物として、配布元（LynnReal-AI）と同じ MiniMax H3 Community License Agreement（LICENSE）で 配布します。追加の条件は付けません。詳細は NOTICE。

• 地域: EU・英国・韓国・米国では、モデルと出力の使用・複製・改変・配布・表示の権利は与えられていません
• 商用: 商用の製品・サービスの年間売上が 2,000 万米ドルを超える場合は、事前に MiniMax の書面許諾（api@minimax.io）が必要です。MiniMax H3 を使う商用の製品・サービスは UI に「MiniMax H3」を目立つように表示します
• 出力: モデルや出力を、MiniMax H3 とその派生以外の AI モデルの改善に使ってはいけません
• AUP: ライセンスの Exhibit A（利用規定）が適用されます