Perflow-Shuai/H3-AR-Few-Step-LoRA

H3 AR Few-Step LoRA

H3 自回归（AR）少步音视频生成 LoRA：每个 chunk 使用 3-step Euler 采样，常规 chunk 约 1.42 秒。

A MiniMax-H3 LoRA for autoregressive, 3-step-per-chunk, joint audio/video generation.

Inference settings

Setting	Value
Base model	MiniMaxAI/MiniMax-H3, revision fa9c8ab1eaa21c8ae25e7e40b83b2e6002f340af
Generation mode	Causal autoregressive (AR), joint video and audio
Denoising	3 Euler steps per chunk
LoRA	Rank 128, alpha 128, scale 1.0
Weights	416 FP32 tensors, native H3 fused layout
Evaluated clip	1344 × 768, 124 frames, 24 fps, approximately 5.17 seconds
Audio	Jointly generated audio, 32 kHz, stereo

Chunk size

For the evaluated first request, the video-latent chunks are [12, 10, 10, 5], corresponding to decoded video-frame chunks [39, 34, 34, 17].

Chunk	Video latent frames	Decoded video frames	Duration at 24 fps
First	12	39	1.625 s
Regular	10	34	1.417 s
Regular	10	34	1.417 s
Final remainder	5	17	0.708 s

Thus, the normal AR chunk size is 10 video latent frames / 34 decoded frames / approximately 1.42 seconds. Three denoising steps apply to each chunk; the four-chunk clip uses twelve student denoising evaluations in total.

Sampler

Use Euler sampling with four sigma endpoints (three evaluations). Start from a 50-point FP32 grid linearly spaced from 1 to 0, select indices [0, 16, 33, 49], and apply:

shifted_sigma = shift * sigma / (1 + (shift - 1) * sigma)
video shift = 12
 audio shift = 3
model time = 1 - shifted_sigma

See inference_config.json for the machine-readable settings. This JSON describes the evaluated configuration; it is not a standalone pipeline implementation.

Download and load

from huggingface_hub import snapshot_download

adapter_dir = snapshot_download(
    repo_id="Perflow-Shuai/H3-AR-Few-Step-LoRA",
    allow_patterns=[
        "adapter_model.safetensors", "adapter_config.json",
        "inference_config.json", "LICENSE", "NOTICE", "SHA256SUMS",
    ],
)

Load the adapter at scale 1.0 on the compatible MiniMax-H3 base using an H3 causal AR runtime. Base weights, the text encoder, video/audio codecs, and the AR inference implementation are separate dependencies.

The native tensor keys use lora_a / lora_b for attn.qkv_proj, attn.out_proj, mlp.fc1, and mlp.fc2 across 50 main blocks and 2 token-refiner blocks. Loading must preserve the fused Q/K/V layout and native FFN gate/value ordering. Generic PEFT or Diffusers load_lora_weights compatibility has not been validated for this format.

The documented geometry covers the first approximately 5.17-second request. Longer autoregressive rollouts have not been validated. Use SHA256SUMS to verify the downloaded files.

License

Powered by MiniMax H3. See LICENSE and NOTICE for the license and attribution.