VQV Signal

RESEARCH WATCH TECHNICAL

X post describes Stanford and Luma AI work that drops the decoder from latent video diffusion

According to a post on X by @Machinedelusion, work from Stanford and Luma AI uses a linear head on a diffusion transformer that predicts pixels directly, instead of going through a decoder. The post reports 23% less warping on Wan2.1 14B video. The figures come from the post's own summary, and the author says they did not check them against the paper. The code is listed as "coming soon."

Source: Scobleizer News Pulse (X) · x.com Found via Robert Scoble’s Scobleizer News Pulse · a post on X, not independently verified Published 2026-11-20T02:45:00+00:00 Detected 2026-11-20T15:02:43+00:00
View original source

According to a post on X by @Machinedelusion, work from Stanford and Luma AI uses a linear head on a diffusion transformer that predicts pixels directly, instead of going through a decoder. The post reports 23% less warping on Wan2.1 14B video. The figures come from the post's own summary, and the author says they did not check them against the paper. The code is listed as "coming soon."

AI-assisted summary based on the listed source.

Machine Delusions highlighted work from Stanford and Luma AI in which a linear head on the diffusion transformer predicts pixels directly instead of going through a decoder, reporting 23% less warping on Wan2.1 14B video, with code "coming soon." It drew about 360 views and 3 bookmarks; the numbers are the post's...

Researchers and developers working on AI video generation may be interested if the approach holds up, since it removes the decoder step from the pipeline. Anyone working with Wan2.1 or similar models should watch for the promised code and the full paper to check the 23% figure. For now, the result rests on a single secondhand summary.

Signal Strength 93% Technical label WATCH Public Interest 54 Category RESEARCH Reader Depth TECHNICAL

Signal Strength reflects source quality, relevance, freshness and evidence. Public Interest helps organize discovery; it is not proof of truth.

Public Interest components
Recognizable Entity Score 73 Practical Impact Score 46 Novelty Interest Score 70 Consequence Score 18 Curiosity Score 32 Shareability Score 65

VQV surfaced this signal because it is recent, relevant to AI Video, connected to Scobleizer News Pulse (X).