PRE-RENDERED · NO GPU IN THIS SPACE

Few-step video on MiniMax H3, decoded without the VAE

DiR generates a clip in 4 iterations instead of 20 steps. DWD reads the pixels straight off the transformer instead of running the VAE decoder. Every clip here is real model output, rendered on one RTX 5090.

Why there is no live demo: the MiniMax H3 Community License does not allow use of H3 or models derived from it in the European Union, the United Kingdom, South Korea or the United States, and a public Space would run it for visitors from anywhere. So this page shows pre-rendered results only. Running the weights yourself is subject to the same licence.

Update, 22 November: pi-H3 is the faster route

The DWD authors released pi-H3, the official 2- and 4-step MiniMax H3 with DWD built in, by Hansheng Chen, Shengqu Cai, Jiaming Song, Gordon Wetzstein, Xinchen Yan and Leonidas Guibas. We measured it on the same RTX 5090, 2 steps with its DWD, through their ComfyUI nodes, with dense attention and with Spark-Attn as in their RTX 50-series workflow.

pi-H3 2-step + DWD, RTX 5090DenseSpark-Attn
768×512, 39 frames (1.6 s)1.39 s1.43 s
1344×768, 39 frames4.98 s5.09 s
768×512, 124 frames (5.2 s)6.49 s4.92 s
1344×768, 124 frames29.6 s12.9 s (2.3×)

Transformer plus DWD decode, after warm-up. Spark stays dense below 12,288 tokens, so short clips gain nothing. On 5-second clips it keeps the scene and motion and changes fine detail. For comparison, DiR with our DWD takes 4.1 s at 768×512 and runs out of memory at 1344×768 × 124 frames. More on the explainer.

Time per 39-frame clip, 768×512, RTX 5090 (DiR, 20 November)

A 39-frame clip is 1.6 s of video at 24 fps, so 1.0 s is faster than real time. The last two rows make a different, softer video than full-resolution DiR, not the same video sooner. DWD's decode alone: 0.26 s against the VAE's 1.8 s.

DWD against the VAE

Same DiR output, three decoders side by side. LPIPS against the VAE on held-out clips: linear head 0.267, with the refiner 0.114.

A held-out portrait. Left: the H3 VAE decode. Middle: the linear DWD head alone. Right: DWD with the refiner.
An isometric scene. Left to right: VAE | DWD head | DWD with the refiner.

The DiR generator

The published DiR checkpoint at 1344×768: 4 iterations for 39 frames, decoded by the H3 VAE.

Fewer tokens: MPEG-Diffusion's two-track sampler

Full-resolution DiR, standard two-track and fast two-track, middle frame
1344×768, left to right: full-resolution DiR, standard two-track, fast two-track (region outlined). Two-track takes its layout and style from half-resolution noise, so it is a different video: here the chef turns into a cartoon.

The v0.2 refiner also learns the two-track region (LPIPS against the VAE 0.065 → 0.041 at 768×512). Two-track stays softer than full resolution; details in issue #7.

Numbers

39 frames, RTX 5090768×5121344×768DWD LPIPS vs VAE
DiR + VAE5.49 s16.77 s–
DiR + DWD4.19 s13.95 s0.100 / 0.108
Two-track fast path + VAE3.08 s7.57 s–
Two-track, DWD on both tracks1.84 s5.07 s0.219 / 0.181
Half res + DWD, pixel ×21.02 s2.88 s0.223 / 0.169

Medians of 2 warm prompts. LPIPS compares DWD with the VAE decode of the same latent; it does not compare one path with another.

Run it yourself

Needs a 32 GB card with no block swap. DWD decodes only the published DiR checkpoint.

git clone -b feat/accel https://github.com/scrya-com/dir-diffusion && cd dir-diffusion
hf download johndpope/dir-diffusion-h3-dwd --local-dir dwd
python3 lot/accel_h3.py check --run johndpope/dir-diffusion-h3 \
  --lora dwd/dwd_lora.pt --head dwd/dwd_head.pt --refiner dwd/dwd_refiner.pt