Sayak Paul’s Post

Post-training in diffusion models is a very under-appreciated topic. So, we're delighted to try to change that at ECCV'26. Announcing a dedicated tutorial for it with the best pack 🔥 We'll cover several tracks & check out the link below to know more! Hila Chefer, Linoy Tsaban

  • No alternative text description for this image

I have identified a fundamental vulnerability in AI models that I have already reported to regulators. I am open to demonstrating this finding under the supervision and control of relevant development teams. For those interested in the full details, I’ve covered this in my last post: https://www.linkedin.com/posts/eduard-pukanych-88335439a_news-ai-artificialintelligence-activity-7466870723056328705-McsF?utm_source=share&utm_medium=member_desktop&rcm=ACoAAGHweOMBMf2WN8WkrW3OlD7Ow-37kJxmlpg

Like
Reply

Post-training for diffusion models feels under-discussed compared to LLMs. This is a useful direction, especially as image and video models move closer to real product workflows.

Like
Reply

Post-training for diffusion models definitely deserves more attention Sounds like an exciting tutorial lineup for ECCV’26

Like
Reply

Post-training being under-appreciated tracks, most of the public attention still goes to base architectures while the preference tuning is what actually shapes what ships. The harder problem you'll probably hit is evaluation: image preference is fuzzier than text, and reward hacking tends to show up as aesthetic mode collapse where everything drifts toward the same overcooked look. Curious whether the tutorial gets into how to catch that before it bakes into the model.

Like
Reply
See more comments

To view or add a comment, sign in

Explore content categories