liran12345/h3-turbo-diffusion

H3 Turbo Diffusion

Full BF16 (16-bit), four-update DMD transformer derived from MiniMax H3 FL2VA for image-to-video generation with synchronized audio.

These are the complete DMD DiT weights, not a rank128/rank256 SVD LoRA, not a LoRA-merged approximation, and not W8A8-quantized weights. The source DMD checkpoint is uploaded without re-serialization or changes to its tensor values. The original checkpoint stores 626 tensors in BF16 and 12 input/output and time-embedding tensors in FP32 (about 66 MiB); these original precisions are preserved. This is an independent derivative, not an official MiniMax release.

Contents

• transformer/: all 34 original safetensors shards, config and index.
• adaln/: the exact DMD AdaLN table for the schedule below, plus its metadata.
• configs/h3_i2va_bf16.yaml: reference configuration for the H3 runtime.
• configs/h3_i2va_w8a8.yaml: configuration for an optional local W8A8 export.
• manifest.json: checkpoint tensor counts, file sizes and SHA256 checksums.
• LICENSE and NOTICE: upstream model terms and derivative attribution.

The transformer weights are approximately 61.73 GiB. 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, tokenizer and other original H3 components are not bundled. Obtain those from MiniMaxAI/MiniMax-H3 and configure the original FL2VA model root separately.

Download

hf download liran12345/h3-turbo-diffusion \
  --local-dir /models/h3-turbo-diffusion

Inference Settings

Use the H3 inference implementation with the DMD Diffusers checkpoint loader and matching H3 architecture support. This is a transformer checkpoint, 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:

sample:
  task: i2va
dit:
  model_path: /models/h3-turbo-diffusion/transformer
  adaln_bake_path: /models/h3-turbo-diffusion/adaln/adaln_i2va_s5_v12_a3.safetensors
  steps: 5
  video_shift: 12.0
  audio_shift: 3.0
  dbcache:
    enabled: false

steps: 5 means five sigma points, giving four denoising updates in the H3 runtime. It is not a five-update schedule. The AdaLN table is specific to these source weights, the i2va task and these shifts. Do not reuse it for a different schedule.

The supplied reference config retains Sparge top-k 0.5, SP8, CPU block offload, quantized/asynchronous communication and a separately exported INT8 video VAE. Thus "BF16" describes the DiT weight/Linear path, not every operation in the pipeline. Edit all /models/ and /work/ paths for your deployment; prompt and first-frame encoding are separate runtime stages. H3 inference code and its dependencies are not included in this model repo.

Optional TurboDiffusion W8A8 Export

For the existing H3/TurboDiffusion integration, users can export W8A8 locally from this full BF16 transformer. No LoRA merge is needed:

cd /path/to/h3-fast-infer
PYTHONPATH=/path/to/TurboDiffusion:. CUDA_VISIBLE_DEVICES=0 \
python example/export_dit_turbo_int8.py \
  -i /models/h3-turbo-diffusion/configs/h3_i2va_w8a8.yaml \
  --base-model-dir /models/h3-turbo-diffusion/transformer \
  -o /models/h3-turbo-diffusion/model_w8a8.safetensors

This requires an H3 runtime checkout containing that exporter and TurboDiffusion built with its CUDA W8A8 operators. Sparge and the VAE dependencies are separate. W8A8 conversion can change generated results; the uploaded transformer itself remains unquantized.

Limitations

The full BF16 DMD model was used as the reference for five 768p I2VA test cases. That is a small functional evaluation, not 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 MIT or 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 and derived 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.