Hansheng Chen 陈涵晟

Blog/Announcement

pi-H3: Video Generation in 2–4 Steps Without a VAE Decoder

Authors
Hansheng Chen1*, Shengqu Cai1, Jiaming Song†, Gordon Wetzstein1, Xinchen Yan†, Leonidas Guibas1
1Stanford University
*Part of work done during an internship at Luma AI †Work done while at Luma AI
Published
Cite
BibTeX
Try the demo Models (coming soon) Code (coming soon)

We announce pi-H3, the fastest version of MiniMax H3 [1] yet. pi-H3 generates video in just 2 or 4 sampling steps and directly predicts pixels at the final step, skipping the video VAE decoder. On a single B200 it produces 5-second HD videos in 3.8 seconds (sub real-time) and 15-second videos in 18.6 seconds, more than twice as fast as Sol-H3 [4].

pi-H3 is an open-source research project. The Hugging Face demo is live today. All model weights and training/inference code will follow in LakonLab [2], with additional ComfyUI support through the ComfyUI-piFlow [3] extension.

  1. Available nowDemo on Hugging Face
  2. Week of Nov 16Model weights and inference code
  3. Later in NovemberTraining code, once we finish cleaning up the codebase

End-to-end generation speed

Measured end to end on a single NVIDIA B200, pi-H3 generates a 5-second 768p video with stereo audio in 3.8 s with 2 steps, and a 10-second video in 9.2 s, both faster than the video plays. The chart below compares complete requests, from text encoding to decoded video and audio, against Sol-H3 [4]. Across 5–15 s videos, pi-H3 is 2.4–2.7× faster with 2 steps and 1.2–1.4× faster with 4 steps. The speedup comes from fewer denoising steps, decoder removal, and FP8 quantization.

  Sol-H3 (official, from its paper) pi-H3 (ours)
GPU 1 NVIDIA GB200 GPU 1 NVIDIA B200 GPU
Sampling steps 4 2 or 4
Attention Sol-Attn [5] on cuDNN BSA Sol-Attn [5] on cuDNN BSA
Linear layers MXFP8 W8A8 in DiT blocks 2–⁠46; boundary blocks, AdaLN, refiners and text encoder in BF16 FP8 W8A8 in all DiT blocks and the text encoder; AdaLN and output layers in BF16
Video output Sol-H3-VAE [4] decoder pixels from the last denoising step (DWD), no VAE decoder
pi-H3 Sol-H3
pi-H3 2-step
3.81 s
pi-H3 4-step
7.3 s
Sol-H3 4-step
9.48 s
0 s 2 s 4 s 6 s 8 s5.2 s of video
Show numbers
Methodpi-H3Sol-H3Total
pi-H3 2-step3.813 s– 3.81 s
pi-H3 4-step7.304 s– 7.3 s
Sol-H3 4-step–9.475 s 9.48 s
End-to-end latency for one 1344×768, 24 fps video with stereo audio (switch the length above the chart). pi-H3 measures the median of seven requests, covering text encoding, all denoising steps, DWD pixel output, RGB conversion, and audio decoding. Model loading, compilation, warm-up and MP4 encoding are excluded. Sol-H3 uses official results from its paper (Table 1), covering the same end-to-end process. The dashed line marks the length of the video itself.

Model latency breakdown

We time each stage of the pipeline separately on a single NVIDIA H100 80GB (batch size 1) and add up the per-stage medians. All methods share the same text encoder, audio decoder and FP8 MiniMax H3 denoiser with Sol-Attn [5], so the differences come from the number of denoising steps and from the video decoder. pi-H3 decodes pixels within its last denoising step (DWD), which removes the video VAE decoder from the pipeline.

Text encoder Denoiser (all steps) Video decoder Audio decoder
2 steps
pi-H3 2-step DWD
6.1 s
6.2 s
PDMD 2-step + H3 VAE
5.8 s4.0 s
9.9 s
4 steps
pi-H3 4-step DWD
12.0 s
12.0 s
Sol-H3 4-step + Sol-H3-VAE
11.7 s
12.6 s
LightX2V 4-step + LightVAE-Lite
11.7 s
13.0 s
0 s 5 s 10 s
Show numbers
MethodText encoderDenoiser (all steps)Video decoderAudio decoderTotal
pi-H3 2-step DWD0.037 s6.125 s–0.032 s 6.19 s
PDMD 2-step + H3 VAE0.037 s5.83 s3.972 s0.032 s 9.87 s
pi-H3 4-step DWD0.037 s11.955 s–0.032 s 12.02 s
Sol-H3 4-step + Sol-H3-VAE0.037 s11.66 s0.88 s0.032 s 12.61 s
LightX2V 4-step + LightVAE-Lite0.037 s11.66 s1.265 s0.032 s 12.99 s
Per-stage latency for one video on a single NVIDIA H100 80GB (switch the workload above the chart). Each bar sums the measured median of every stage. Model loading, compilation, and warm-up are excluded. pi-H3’s last denoising step costs slightly more than a regular step because it also produces pixels, replacing the video decoder. Hover a segment for its time.

Uncompromised fidelity at 2–4 steps

pi-H3 offers superior motion, texture, and audio fidelity compared to other few-step MiniMax H3 models, while maintaining structural coherence. The following comparisons show results from the same prompts and Gaussian noise for each model, without seed cherry-picking.

At 4 steps, pi-H3 delivers the best overall quality among the 4–8-step MiniMax H3 models we compared: LightX2V Turbo [6], FastH3 Dense/DataFree [7], and larryvrh’s H3-Turbo v4 [8]. It generates more dynamic, vivid movement, whereas the other models tend toward reduced, mechanical motion. Frames are sharper overall, with richer texture and finer detail. On the audio side, pi-H3 produces high-fidelity sound with full bass and crisp speech, while the other models sometimes generate broken audio or lose the bass.

Few-step MiniMax H3 models with 4–8 steps. Every model gets the same text embeddings and Gaussian noise and runs its official schedule. All evaluations use the same FP8 Sol-Attn inference stack with the native H3 VAE decoder. pi-H3 is outlined in blue. Use the speaker button to hear a clip, and double-click a clip (or use its expand button) to view it fullscreen.

The same strengths carry over to the 2-step setting: pi-H3 delivers more dynamic, realistic visuals and higher-fidelity audio compared to the 3-step TaoMate-H3 [9] and 2-step PDMD [10]. At this quality level, 2-step generation is already sufficient for simpler scenes.

Few-step MiniMax H3 models with 2–3 steps. Every model gets the same text embeddings and Gaussian noise and runs its official schedule. All evaluations use the same FP8 Sol-Attn inference stack with the native H3 VAE decoder. pi-H3 is outlined in blue. Use the speaker button to hear a clip, and double-click a clip (or use its expand button) to view it fullscreen.

Direct pixel output without video decoding

pi-H3 adopts DWD, a novel method that produces pixels directly from the last denoising step, bypassing the heavy VAE decoder. For few-step generation, this yields a massive speedup. The following comparison shows pi-H3 DWD output alongside non-DWD pi-H3 latents decoded by the native H3 VAE, Sol-H3-VAE [4], and LightVAE-Lite [11, 6].

Decoding methods. DWD generates video without a VAE decoder, from the same prompt and noise as the others. H3 VAE, Sol-H3-VAE and LightVAE-Lite decode the same pi-H3 latents, so their frames line up pixel for pixel. Drag the slider to compare any two; switch to Grid to see all four.

How it works

pi-H3 combines two novel techniques: pi-Flow v2 and DWD.

pi-Flow v2 is a full upgrade of pi-Flow [12] for data-free few-step distillation. It retains pi-Flow’s strength of producing rich texture while substantially improving structural coherence, especially in the 2-step setting. It also simplifies inference: pi-Flow v2 models typically produce straight trajectories, removing the complexity of pi-Flow’s substep integration.

DWD (Decoding Without a Decoder) removes the video decoder from the pipeline entirely. Instead of passing final latents through a separate VAE decoder network, DWD re-uses the denoiser itself to predict clean pixels at the last step. DWD on MiniMax H3 is still experimental: H3 uses an unusually large compression ratio (4×32×32 3-channel pixels per token), and we have not yet had enough compute to fully optimize for it, leaving room for further improvement.

More details will be revealed in upcoming releases.

All videos below are generated by pi-H3 with 2 or 4 sampling steps, without the video VAE decoder, at up to 768p with stereo audio. Open a gallery, then click any video to watch it with sound and read its full prompt.

Text to video · 2 steps12 videos
Text to video · 4 steps10 videos
Image to video3 videos
First and last frame to video2 videos

Try it

If the demo doesn't load here, open it on Hugging Face.

Citation

Hansheng Chen, Shengqu Cai, Jiaming Song, Gordon Wetzstein, Xinchen Yan, Leonidas Guibas. “pi-H3: Video Generation in 2–4 Steps Without a VAE Decoder.” November 2026. https://hanshengchen.com/blog/pi-h3/

BibTeX 
@misc{chen2026pih3,
  title        = {{pi-H3: Video Generation in 2–4 Steps Without a VAE Decoder}},
  author       = {Hansheng Chen and Shengqu Cai and Jiaming Song and Gordon Wetzstein and Xinchen Yan and Leonidas Guibas},
  year         = {2026},
  month        = nov,
  howpublished = {\url{https://hanshengchen.com/blog/pi-h3/}}
}

References

  1. [1] MiniMax. MiniMax H3: An Open Model Breaking the Boundaries Between Tasks and Modalities. MiniMax Research Blog, 2026. www.minimax.io/blog/minimax-h3
  2. [2] Hansheng Chen. LakonLab. GitHub repository, 2026. github.com/Lakonik/LakonLab
  3. [3] Hansheng Chen. ComfyUI-piFlow. GitHub repository, 2026. github.com/Lakonik/ComfyUI-piFlow
  4. [4] Yitong Li, Jincheng Yu, Junsong Chen, Haopeng Li, Shuchen Xue, Haozhe Liu, Ping Luo, Song Han, and Enze Xie. Sol-H3: Recursive Self-Improvement for MiniMax-H3 Inference Acceleration on Sol-Engine across Cloud and Edge. 2026. arxiv.org/abs/2609.35110
  5. [5] Haopeng Li, Yitong Li, Junsong Chen, Tian Ye, Haozhe Liu, Jincheng Yu, Duomin Wang, Ruihua Zhang, Zeke Xie, Enze Xie, and Song Han. Sol-Attn: Accelerating Video Generation Inference via On-the-Fly Attention Sparsification. 2026. arxiv.org/abs/2607.24027
  6. [6] LightX2V Contributors. LightX2V: Light Video Generation Inference Framework. GitHub repository, 2025. github.com/ModelTC/lightx2v
  7. [7] FastVideo Team. FastH3 Preview v1: Four Open-Weight H3 Models in Four Steps. Blog post, 2026. haoailab.com/blogs/fasth3-preview/
  8. [8] larryvrh. MiniMax-H3 Turbo Lora. Hugging Face model repository, 2026. huggingface.co/larryvrh/MiniMax-H3-Turbo-Lora
  9. [9] Qijun Gan, Chenwei Zhang, Meiguang Jin, Junfeng Ma, and Qiu Shen. TaoMate: Anchor-Guided Memory Bridging Evolving and Reference States for Real-Time Audio-Video Digital Human Generation. 2026. arxiv.org/abs/2607.24359
  10. [10] Zimo Wang, Junkun Yuan, Angtian Wang, Haotian Yang, Canyu Zhang, Siyuan Yuan, Xingchang Huang, Bo Liu, Yizhi Wang, Yiding Yang, Chongyang Ma, and Gordon Guocheng Qian. PDMD: Projected Distribution Matching Distillation for Video Diffusion Models. 2026. arxiv.org/abs/2609.35768
  11. [11] Shiqiao Gu, Yang Yong, Fengjun Zhong, Ruihao Gong, Chengtao Lv, Xiangyu Fan, Lei Yang, and Xianglong Liu. LightVAE: Towards Compact and Efficient Video Autoencoders. 2026. arxiv.org/abs/2610.14726
  12. [12] Hansheng Chen, Kai Zhang, Hao Tan, Leonidas Guibas, Gordon Wetzstein, and Sai Bi. pi-Flow: Policy-Based Few-Step Generation via Imitation Distillation. ICLR, 2026. arxiv.org/abs/2510.14974