GenCeption Matches Specialist Vision Models with Far Less Data

Every so often a paper argues that a whole subfield's specialist toolkit might be a passing phase, and this week's candidate arrives in a new arXiv preprint from Kaiming He, Andrew Zisserman, Joao Carreira and colleagues. Their claim is that a pre-trained text-to-video generation model, wrapped in a system they call GenCeption, can handle a spread of classic vision tasks competitively with the dedicated specialists built for each of them, and can reach that bar with dramatically less task-specific data.

The specialists it takes on are DepthAnything3, SAM3, and D4RT, across depth estimation, segmentation, and 3D keypoint prediction. GenCeption reportedly holds its own against those systems while using, per the paper, '7 to 500 less training data'. That is the headline number, and it is why the result is worth reading even if you do not work in vision. The same generative-pretraining recipe that reshaped NLP appears to be starting to work on the pixel side, with a video model playing the role that a language model played for text.

The other detail worth pulling out is generalization. The authors report that GenCeption models trained only on synthetic human videos transfer to real-world footage and to categories the training set never included, including animals and robots. If that holds up in independent hands, it changes the cost calculus for anyone building perception systems, because the expensive part today is curating labeled data per task and per domain.

The honest caveats: the paper positions video generative pretraining above alternatives like V-JEPA and Video MAE under its own experimental conditions, which is not the same as a settled result once other teams reproduce it. The efficiency multiplier is reported as a wide range rather than a tight one, and the write-up does not give you compute cost, inference latency, or a sense of where the specialists still win on hard edge cases. Take the numbers as directional, not final.

If the direction holds, the winners are the labs and startups that already invested in large video-generation stacks. The exposed side is anyone whose moat was a single-purpose vision model trained on a proprietary labeled corpus, because the terms of that competition change quickly once a general pretraining backbone can be steered by text to do the same job.