DeepMind: “World models exist within current video generation models”
DeepMind developed GenCeption: A generalist model based on Wan2.1 video generator outperforms task-specific models trained on millions of videos by up to 500x less data.
New research from Google DeepMind presents GenCeption: a model that takes the Wan2.1 video generation model and repurposes it for general computer vision problems. Rather than training dedicated models for each task (depth estimation, object segmentation, 3D pose reconstruction), it is found that Wan2.1 can also be used for these tasks.
GenCeption is pre-trained on 7500 synthetic videos and achieves comparable or superior performance to state-of-the-art models which require 7–500x more data. In particular, it matches DepthAnything3 in depth estimation performance and surpasses NormalCrafter in surface normal prediction. Surprisingly, when trained on single human subjects, GenCeption still generalizes effectively to multi-person scenes, animals, and robots.
Video generators build an internal model of physics and space during training. DeepMind proposes using this as a ready-made backbone instead of developing dozens of narrow specialists. One model — many computer vision tasks, without costly datasets.
Source: the-decoder.com
Free course
Stop reading about AI — start building with it
The free Claude Code course: your first site, tool or game — no coding. No upsells, no cross-sells — nothing to buy here.
Start free →
Author
Evgenii Arsentev
PhD · Chief Executive Officer, digital health
Articles · Latest articles