Log inSign up
user avatar
Or Patashnik
@OPatashnik
Assistant Professor @ Tel-Aviv University
orpatashnik.github.io
Joined May 2019
549
Following
1,662
Followers
RepliesMedia
  • user avatar
    Or Patashnik
    @OPatashnik
    Jun 25
    Reference-conditioned diffusion models rely on dense reference token grids. Are they really necessary? Surprisingly, dropping 80–90% of the reference tokens barely affects quality. A little fine-tuning is enough to recover nearly the same quality at a fraction of the cost.
    user avatar
    Rishubh Parihar
    @RishubhParihar
    Jun 25
    🌟🚀 Excited to share our latest work: "Keep The Essentials: Efficient Reference Conditioned Generation via Token Dropping"! TL;DR: Stop wasting compute on redundant tokens! We introduce SparseContext that drops reference tokens for speeding up reference-based image generation⚡
    Image
    00:00
  • user avatar
    Or Patashnik
    @OPatashnik
    Jun 22
    We find a shortcut in training reference-conditioned audio flow matching: at low noise levels, the model assigns speakers based on acoustic similarity rather than text. Biasing timestep sampling toward higher noise removes this shortcut and restores text-driven speaker assignment
    user avatar
    Mickey Finkelson
    @mikikiFin
    Jun 22
    1/7 New paper 🧵 ScenA generates a multi-speaker audio scene: overlapping speech, laughter, real room noise, from a text description and a few reference voices. When trained in the obvious way, it ignores the text and decides who speaks on its own. We found out why and fixed it.
    Image
  • user avatar
    Or Patashnik
    @OPatashnik
    Apr 27
    Really excited about this one! We turn trade-offs between multiple rewards during diffusion model post-training into a controllable inference-time slider. A single post-training run yields a model that can smoothly traverse the Pareto frontier.
    user avatar
    Shelly Golan
    @Shelly_Golan1
    Apr 25
    1/7 When rewards conflict, what should RL post-training of diffusion models optimize? In visual generation, objectives are often in tension: Prompt adherence can conflict with source preservation. Photorealism can conflict with stylization. In our new paper, ParetoSlider, we
    Image
    00:00
  • user avatar
    Or Patashnik
    @OPatashnik
    Apr 11
    We analyzed CFG in diffusion editing models and discovered how to adapt the guidance mechanism to achieve continuous, fine-grained control over edit strength. Excited to present our recent work which was accepted to SIGGRAPH 2026! 🥳✨