Pelican riding a bike with both legs in front: Qwen 4 Max caught by an occlusion bug
A generated video of a pelican riding a bike made Qwen 4 Max the focus of discussion: environmental interaction was praised, while occlusion logic was picked apart. The comments laid the failure point out clearly.
Kaka posted a video generated by Qwen 4 Max on X. The original words were: Is there a bit of a problem with Qwen 4 Max's occlusion logic—how come it drew both of the pelican's legs in front? Other than that, environmental interaction and so on are very good, and the aesthetics finally aren't the same old cookie-cutter GPT style.
Original post link: Kaka's tweet on X. When D.Life reposted it, they wrote only two words: 'So smooth.'
Occlusion logic sounds like a minor issue, but in video it is fatal. When riding a bike, which leg is in front and which leg is blocked by the bike body depends on body structure and camera angle. Both legs ending up in front shows that the model can make it look right in a single frame, but once it gets to continuous frames, the spatial relationships start getting tangled. Ordinary users see a pelican riding awkwardly; people working on video generation see that geometric constraints and spatiotemporal consistency still aren't passing muster.
The comment section did not let this detail go. gg ff said, 'So rock 'n' roll'; Panda Burning Incense added, 'Even a pelican riding a bike has to obey traffic rules'; Chen Jiefu thought it 'looks like a product from a previous era'; Alexa Wang noticed 'the riding looks a bit difficult'; and Patrick Old-school used 'too bumpy.' Kaka replied to Alexa Wang: 'It's having a hard time making it too.'
What's interesting about this video is that people didn't just judge whether it looked real. Some focused on environmental interaction, some on aesthetic style, and some went straight for the occlusion relationship. Users are starting to review videos on physical consistency, with standards more specific than 'can it generate?' This time, what Qwen 4 Max got praised for and where it failed are both very clear: the style is no longer uniformly GPT-style, and environmental interaction has improved; the occlusion bug is also intuitive enough that the position of one leg can break the immersion.
Zero Drift said in a reply, 'The whole internet is laughing at Qwen4 max,' and D.Life replied, 'Too funny.' For ordinary users, this is a funny video of a pelican riding a bike; for model builders, it's a free test case. Now that video generation has come this far, grand narratives are less persuasive than one pelican's leg.
发布时间: 2026-11-11 15:26