Commentary: Video models lag behind LLMs in post-training progress

AccBalanced · x · 2026-08-29

Arankomatsuzaki observes that the failure mode of generating 'no text, no logos' is common across image and video generation. He hypothesizes that LLMs are simply further along in post-training and RL compared to image models, which in turn are ahead of video models. He also notes that Seedance is further ahead in video generation than he had previously assumed.

Original post →

More from Models

Models channel →