TurboH3-FL2VA-I2VA

A four-update DMD transformer derived from MiniMax H3 FL2VA for image-to-video generation with synchronized audio, published for the TurboDiffusion H3 integration.

The repository provides:

• the complete BF16 DMD transformer, and
• a W8A8 (INT8) export of it for TurboDiffusion's native quantized kernels.

This is an independent derivative, not an official MiniMax release. It was originally published at liran12345/h3-turbo-diffusion and moved here with the author's consent. All transformer shards and the AdaLN table are byte-identical to that release.

Contents

Path	Size	Description
transformer/	61.73 GiB	All 34 DMD safetensors shards, config and index
w8a8/model_w8a8.safetensors	20.36 GiB	W8A8 export of transformer/ (format h3_fl2va_turbo_w8a8_128)
adaln/	102 MiB	AdaLN modulation table for the published schedule, with metadata
manifest.json	—	Sizes and SHA256 checksums for every file above
LICENSE, NOTICE	—	Upstream model terms and derivative attribution

The BF16 checkpoint stores 626 tensors in BF16 and 12 input/output and time-embedding tensors in FP32 (about 66 MiB). These original precisions are preserved. The index metadata is normalized to omit internal storage locations; its tensor-to-shard mapping and every shard's bytes are unchanged.

Text encoder, video/audio VAEs, processor and tokenizer are not bundled. The downloader below fetches them from a pinned revision of MiniMaxAI/MiniMax-H3.

Download

From a TurboDiffusion checkout containing the H3 integration:

python -m turbodiffusion.inference.download_h3 \
  --dit-dir /models/TurboH3-FL2VA-I2VA \
  --components-dir /models/MiniMax-H3

The downloader pins both repositories to immutable revisions and verifies every selected file of this repository against manifest.json (size and SHA256). This checks release consistency, not a cryptographic signature. Pass --skip-w8a8 to fetch only the BF16 weights and export W8A8 locally.

W8A8 Weights

w8a8/model_w8a8.safetensors quantizes the 200 main-block QKV, attention output and MLP projections with TurboDiffusion's existing Int8Linear (128x128 weight blocks). It keeps the original FP32 input/output/timestep tensors and RoPE frequencies. AdaLN weights are omitted, so the file must be used together with adaln/.

Export provenance is recorded in w8a8/model_w8a8.json: TurboDiffusion commit bfe1067819d09faa6d9ce0eee47de6cb520205d9 (PR #144), PyTorch 2.11.0+cu128, CUDA 12.8, on RTX 5090.

The INT8 values depend on TurboDiffusion's quantization kernel and file format. If a later TurboDiffusion version changes either, regenerate the file from transformer/ with its exporter:

CUDA_VISIBLE_DEVICES=0 python -m turbodiffusion.inference.export_h3_w8a8 \
  --model-path /models/TurboH3-FL2VA-I2VA/transformer \
  --output /models/TurboH3-FL2VA-I2VA/w8a8/model_w8a8.safetensors

W8A8 conversion can change generated results. The BF16 transformer remains the reference.

Inference Settings

Use the TurboDiffusion H3 implementation and its docs/h3_i2va.md. These are transformer weights, not a standalone DiffusionPipeline.from_pretrained bundle. The MiniMaxH3Transformer3DModel config does not imply that every Diffusers release includes that model class.

The verified schedule is i2va / steps=5 / video_shift=12 / audio_shift=3. Five sigma points give four denoising updates. The AdaLN table is specific to these weights, the task and the shifts; do not reuse it for a different schedule. It was generated with the source checkpoint's FP32 timestep MLP, recorded as "timestep_mlp_dtype": "float32" in its JSON and embedded metadata. The table itself is stored in BF16.

SpargeAttn is an external dependency. The video decoder applies native W8A8 projections to the public VAE weights at startup; no separate VAE export is required. Audio VAE decoding remains FP32.

Limitations

The BF16 DMD model was evaluated on five 768p I2VA test cases. That is a small functional evaluation, not a comprehensive quality, safety or audio evaluation. This release does not claim lossless parity with the original many-step FL2VA model or Wan-specific speedups.

License

The weights are governed by the MiniMax H3 Community License Agreement, not TurboDiffusion's code license. The license includes territorial, use and commercial conditions; its excluded territories include the EU, UK, Republic of Korea and USA. Review the full LICENSE and obtain any required permissions before downloading, using or redistributing. The upstream acceptable-use policy is included in that agreement.

The DMD transformer, the W8A8 export and the AdaLN table are modifications of the original H3 works. See NOTICE for attribution. Generated outputs must be identified as AI-generated as required by the upstream terms.