生成式影片與推論加速
FastH3 Compresses MiniMax H3 to Four DiT Calls, Generating 15 Seconds of Video in 13 Seconds on Eight B200 GPUs
FastVideo combines DMD2 distillation with 90% sparse attention to reduce the base model’s 49 DiT calls to four, while releasing the full weights and LoRA. Its real-time generation figures depend on eight B200 GPUs, specialized VSA kernels, and fully warmed-up tests; a single GPU and mainstream consumer GPUs have yet to achieve comparable speeds.
The FastVideo team, led by UC San Diego’s Hao AI Lab, has released FastH3 Preview v1, a four-step distilled version of MiniMax H3’s text-to-video-and-audio model. The base H3 model’s 33B-class audiovisual diffusion transformer requires 49 passes for each output. FastH3 uses DMD2 distribution matching distillation to reduce the denoising process to four DiT calls, then combines it with Video Sparse Attention (VSA-H3), enabling the recommended checkpoint to maintain approximately 90% sparsity in attention computation. The team has released the full weights, a pre-extracted LoRA, a dense-attention baseline, and FastVideo inference code.
In official tests at 1344×768 and 24 FPS with stereo audio, generating 15 seconds of content took 47.2 seconds on one B200, 15.5 seconds on four B200s, and 12.88 seconds on eight B200s. On a single GPU, this represents a 14.38× speedup over the base H3 model using dense FlashAttention 4. However, because no base-model benchmark was provided for an equivalent eight-GPU configuration, the team did not claim an eight-GPU speedup multiplier. The figures are medians from three requests after a full warm-up and include encoding, denoising, video and audio decoding, muxing, and file output, but exclude model loading and compilation time.
The acceleration comes from two separable components: few-step distillation reduces the number of transformer forward passes, while sparse kernels reduce the attention workload at each step. This shows that video inference optimization need not rely solely on quantization or tensor parallelism; post-training can also directly alter sampling cost. However, the recommended VSA checkpoint depends on the tile-64 VSA-H3 backend and specialized kernels. Users cannot simply replace them with dense attention and expect the same quality and speed. The easier-to-deploy dense distilled version still takes 91.3 seconds to generate a 15-second video on a single GPU.
Preview v1 currently supports only text-to-video-and-audio generation; first-and-last-frame conditioning and reference-image modes have not yet been distilled. The team says training consumed more than 1,000 B200 GPU-hours, but the complete training recipe is still marked as forthcoming. The weights also remain subject to the MiniMax H3 Community License rather than a standard Apache or MIT license. Key areas to watch next include benchmarks on RTX GPUs, DGX Spark, and Apple MLX; cold-start latency; peak memory usage; and independent evaluations of how sparse distillation affects complex motion, fine detail, and audio-video synchronization.