Lakonik/pi-H3 on Hugging Face ships 2-step and 4-step adapters that make MiniMax H3 video generation dramatically faster by combining pi-Flow v2 few-step distillation with DWD (Decoding Without a Decoder) — predicting pixels on the last denoising step instead of running the full video VAE decoder. The official pi-H3 blog from Stanford-affiliated authors reports 3.8 seconds to generate a 5-second 768p clip with stereo audio on one NVIDIA B200, faster than real time playback.
Weights, LakonLab inference code, and a Hugging Face demo are live as of November 21, 2026 (pi-H3 1.0); ComfyUI workflows ship through ComfyUI-piFlow.
TL;DR
| Item | Detail |
|---|---|
| Base model | multimodalart/MiniMax-H3-Pruned plus Lakonik adapters |
| Steps | 2 or 4 (DWD adapters recommended on the card) |
| Speed claim | 3.8 s for 5 s 768p on B200 (2-step, authors' median) |
| Code | LakonLab v0.3, diffusers ComponentsManager |
| License | MiniMax H3 Community License (model derivative) |
| Paper | arXiv:2611.22287 for DWD |
pi-H3 few-step MiniMax H3 video generation benchmark chart without VAE decoder
Image: Lakonik
Why pi-H3 matters after a week of frontier video launches
November 2026 already brought Apple FastFrames-L interactive text-to-video and several closed flagship video models. MiniMax H3 itself is a multimodal open-weights line with tight audio-video coupling — expensive to decode when every clip runs dozens of steps plus a VAE.
pi-H3 attacks the latency stack, not the creative brief:
- Fewer steps via pi-Flow v2 distillation (successor to pi-Flow ICLR 2026 work).
- No separate video decoder pass via DWD, documented in Decoding Without a Decoder.
- FP8 linears and Sol-Attn sparse attention kernels already used in competing H3 turbo stacks.
If you are comparing against Sol-H3 or community LightX2V ports, pi-H3's pitch is end-to-end wall clock, not a leaderboard still frame.
What shipped on Hugging Face
The model card lists four adapter variants:
| Adapter folder | Steps | Output path |
|---|---|---|
pi_h3_2step_dwd_adapter | 2 | DWD pixels (recommended) |
pi_h3_4step_dwd_adapter | 4 | DWD pixels |
pi_h3_2step_adapter | 2 | Classic VAE decoder |
pi_h3_4step_adapter | 4 | Classic VAE decoder |
INT8 ConvRot builds exist for ComfyUI only. All adapters support text, first-frame, and last-frame conditioning.
Release notes on the blog:
- Nov 10 — public demo (pi-H3 0.9).
- Nov 21 — pi-H3 1.0 weights, inference code, ComfyUI; improved 2-step coherence and fine motion vs 0.9.
Downloads on Hugging Face exceeded 9,000 last month when we filed this post — a traction signal for an adapter repo, not a full foundation train-from-scratch release.
How DWD removes the VAE decoder
DWD pixel output compared with H3 VAE and Sol-H3-VAE decoding on shared latents
Image: Hansheng Chen
Classic video diffusion ends with latents → VAE decode → RGB. H3 tokens are unusually fat: each covers 4×32×32×3 = 12,288 pixel values, more than twice the DiT hidden size (5,376). A naive linear DWD head blurs frames.
pi-H3 1.0 replaces the convolutional refiner from early DWD experiments with a windowed transformer refiner using 8 of 24 blocks from the Sol-H3-VAE architecture, tuned end-to-end. That keeps pixels sharp without paying the full decoder cost every request.
The blog's decoder comparison slider shows DWD output against latents decoded by native H3 VAE, Sol-H3-VAE, and LightVAE-Lite — same noise, same prompt, pixel-aligned for non-DWD paths.
Reported speed numbers (and what they include)
pi-H3 end-to-end latency versus Sol-H3 on a single B200 GPU
Image: Lakonik
Authors measure median of seven full requests on B200: text encoder, all denoising steps, DWD pixel output, RGB conversion, and audio decode. They exclude model load, compile warm-up, and MP4 muxing — fair for apples-to-apples research tables, conservative for product SLAs.
5-second 1344×768 @ 24 fps with stereo audio:
| Method | Total time (blog table) |
|---|---|
| pi-H3 2-step DWD | 3.81 s |
| pi-H3 4-step DWD | 7.30 s |
| Sol-H3 4-step + VAE | 9.48 s |
At 10 s length, 2-step pi-H3 reports 9.16 s total vs 24.31 s for Sol-H3 4-step paper numbers — the gap widens with duration because decoder time and attention dominate long clips.
On H100 80GB, a staged breakdown shows video decoder time vanishing for DWD: pi-H3 2-step DWD totals ~6.2 s on 768p·5s workloads vs ~9.9 s for PDMD 2-step + H3 VAE because ~4.0 s of VAE decode disappears.
Quality claims at 2–4 steps
Authors compare matched prompts and Gaussian noise against other few-step H3 ports:
- 4-step: LightX2V Turbo, FastH3, larryvrh H3-Turbo v4 — pi-H3 claims richer motion, sharper texture, and fuller audio bass than mechanical motion from competing few-step models.
- 2-step: beats 3-step TaoMate-H3 and 2-step PDMD on motion and audio fidelity for simpler scenes.
Treat subjective galleries as author-reported; run your own prompts if you ship customer-facing video.
How to run pi-H3 locally (diffusers path)
Install LakonLab v0.3 per the pi-H3 README. The card's Python sketch:
- Load
MiniMaxH3DWDBlockspipeline onmultimodalart/MiniMax-H3-Pruned. load_lakonlab_adapter(pipe, 'Lakonik/pi-H3', subfolder=...)forpi_h3_4step_dwd_adapter(or 2-step).- Optional
quantize_frozen_linearson transformer and text encoder for FP8 W8A8. - Set Sol sparse attention after block 2 if
lakonlab[sol]is installed. - Call
pipe(..., output=['dwd_pixels', 'audio', 'sampling_rate'])thenRGBColorEncoder.decode.
Default example prompt on the card: "A red fox trotting through a snowy pine forest…" at 768×1344, 243 frames (~10 s at 24 fps with H3's 17n+5 frame rounding).
For image-to-video, pass image= or last_image=; the blog gallery includes first/last frame conditioning clips.
ComfyUI and attention accelerators
ComfyUI-piFlow publishes workflows and INT8 ConvRot links. The blog's Complementary accelerations section notes pi-H3 timings already use Sol-Attn; Spark-Attn is another training-free sparse path compatible with Comfy graphs.
Longer term, hybrid linear attention ideas like Video DeltaNet target livestream-length video — orthogonal to pi-H3's step reduction, but combinable in principle.
License and commercial use
The Hugging Face License section is explicit: pi-H3 adapters are model derivatives of MiniMax H3 under the MiniMax H3 Community License Agreement. Read NOTICE on the repo before you ship a paid product.
That is stricter than a plain Apache tool repo. If you already run MiniMax H3 open weights under community terms, pi-H3 inherits the same family rules.
How pi-H3 compares to other open video stacks we cover
| Axis | pi-H3 | Apple FastFrames-L | MBZUAI PAN 2 |
|---|---|---|---|
| Goal | Faster H3 sampling | Interactive 480p control | World model reasoning |
| Steps | 2–4 | Varies | Research sim rollout |
| Decoder | Skipped (DWD) | VAE path | N/A (simulation) |
| Weights | HF adapters + pruned H3 base | HF open weights | Open 29B (separate post) |
Builders picking a stack should match latency, license, and audio coupling — not parameter counts alone.
Hardware and ops checklist
- GPU: Authors benchmark B200 and H100 80GB; consumer cards may need shorter clips or 480p settings.
- Attention: DWD pixel refiner uses variable-length windows — README sets
varlenbackend ondwd_pixel_refiner. - Memory: FP8 quantization excludes norm_out, proj_out, dwd_linear_head, and
lm_headfrom quantization passes on the card snippet. - Frame math: H3 fixes 24 fps; frame counts follow
17n + 5rounding noted in the sample.
Research lineage and citations
Primary citation from the blog:
Chen et al., pi-H3: Video Generation in 2–4 Steps Without a VAE Decoder, November 2026, hanshengchen.com/blog/pi-h3
DWD paper: arXiv:2611.22287. Affiliations list Stanford University with internship work at Luma AI called out on the card.
What to watch next
Authors promise training code after cleanup. pi-Flow v2 details are teased as a fuller release. For production teams, the actionable milestone is whether 1.0 adapters hold up on your brand motion and lip-sync tests — run the HF demo before you reserve GPU weeks. Start with 4-step DWD if you need safer motion, then try 2-step once prompts and conditioning images are stable. Save raw dwd_pixels tensors if you plan to fine-tune color grading; re-encoding from MP4 throws away headroom.
Troubleshooting common pi-H3 setup errors
Readers hitting install issues usually land in one of four buckets:
Missing LakonLab extras. Sol sparse attention after transformer block 2 requires lakonlab[sol] in the environment the README describes. Without it, remove the processor swap loop or stay on dense attention — expect slower runs, not broken outputs.
Wrong adapter subfolder. The DWD adapters (pi_h3_2step_dwd_adapter) differ from VAE-path adapters (pi_h3_2step_adapter). Mixing subfolder with output=['dwd_pixels'] triggers shape errors at the color encoder.
Frame count rounding. H3 expects num_frames aligned to 17n + 5 at 24 fps. Off-by-one frame counts can fail silently in schedulers tuned for H3's training contract.
License scope. Commercial teams sometimes assume adapter weights are MIT because LakonLab is open source. The card ties derivatives to MiniMax H3 Community License — legal review belongs before customer-facing APIs.
Teams comparing cloud APIs against self-hosted H3 should budget encode plus denoise plus DWD as one latency line item. Managed video APIs hide that breakdown; pi-H3's published tables are useful when you argue for on-prem B200 pools versus per-second SaaS bills for marketing clips.
When you log experiments, store adapter version (0.9 demo vs 1.0 weights), step count, MAX_RAW_T, and whether Sol-Attn was enabled. Few-step models drift with scheduler knobs; reproducibility matters more than a single flashy render.
