AI Science Tracker
EN

Video diffusion models were sensitive at similar denoising steps across prompts

University of Chinese Academy of Sciences; Institute of Automation, Chinese Academy of Sciences; KlingAI Research; Peking University · 21 Oct 2026 · computer science

Discovery

Video diffusion models were sensitive at similar denoising steps across prompts

Source check: Source checked. Not independent scientific confirmation.

Finding: Researchers reduced attention in one denoising window at a time. Within each of four pretrained video models, the windows that most changed the final video were similar across eight prompts. Median rank correlations against the other prompts ranged from 0.95 to 1.00.

Technology: Wan2.1, Wan2.2, HunyuanVideo

Publication: Preprint · Test context: direct test. Catalog categories, not an evidence grade.

AI contribution
Researchers selectively altered attention in trained video diffusion models and measured how the resulting videos departed from the dense model output.
Test and result
The experiment compared final-video error with dense attention and ranked ten windows separately for each prompt, holding each prompt out from the comparison rank.
Limitation / still unknown
The result concerns within-model rankings for a small set of prompts and the tested attention perturbation. It does not establish that the same window order applies to other models or sampling settings.
Evidence sources and annotation context

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

Evidence sources: https://arxiv.org/abs/2610.29441 · https://arxiv.org/html/2610.29441v1

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: paper. Paper: linked separately.

Review date: 21 Nov 2026. Review scope: arXiv v1 front matter, §3 attention sensitivity, Figure 2, Appendix D.1 leave-one-out tests.

No follow-up recorded; this is not a claim that no follow-up exists. Follow-up review scope and date are not recorded. No source-supported relationships recorded.

Editorial assessment: Impact 2/5 · Validation 3/5 · Novelty 2/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Open in the full catalog for filters, charts and related discoveries. Browse computer science discoveries.

Related catalog entries

Links use documented relationships where available, plus shared catalog context. A shared topic or model family does not establish a scientific dependency.