ALLOTInteractive viewer

Interactive results

Compare every measured output.

Explore three settings and all 32 paired prompts. Switch among Dense, Default, and ALLOT, step through matching frames, inspect the same crop, and examine the solved layer–step allocation.

01 / SEE THE DIFFERENCE

Default and ALLOT use the same selection strategy and kernel.

Video comparison

This clip · Default PSNR—
This clip · ALLOT PSNR—
Fidelity gain on this clip—
Configuration generation time—Default → ALLOT · measured on A800
Default
+ ALLOT
Loading the paired videos…
DefaultALLOTDrag to compare the same frame
01 / 81
PSNR measures fidelity to the dense result for this prompt and seed.
Explore the layer–step allocation

02 / EXPLORE THE ALLOCATION

A density for
every layer and step.

ALLOT shares one allocation across prompts. Explore the table used in the selected setting: darker cells retain more attention blocks.

—cells use dense attention
50 × 30denoising steps × layers

This is the denoising schedule, separate from the video's playback timeline.

XAttention · 480pStep 00 · Layer 00 · 100%
5% retained
100% · dense
Step 00 · density of each layerRange 100–100%

Layers 0–29 · move the step slider or select a heatmap column.

Mean density 100%
How to read the comparisons

All 32 evaluation prompts are available in each of three Wan2.1-1.3B settings. Each setting opens at the prompt nearest its median paired PSNR gain.

Citation

BibTeX

@misc{xia2026allot,
  title = {Sparsity Should Be ALLOTted by Output: Training-Free Allocation of Block-Sparse Attention for Video Diffusion Transformers},
  author = {Xia, Tianze and Zhou, Xiaoling and Lan, Ruya and Liao, Yiqiao and Song, Chengru and Xiang, Shiming},
  year = {2026}
}