Other

new paper: iREPA diffusion models are a renderer of their underlying representations. with this new...

Original title:new paper: iREPA diffusion models are a renderer of their underlying representations. with this new...

AI summary

New paper: iREPA diffusion models are a renderer of their underlying representations. With this new setup, we can gain clearer insight into what these representations really mean. Jas began a spontaneous exploration, and we learned a lot over the past three months. ps. This is also our small experiment with a new kind of online 'water cooler effect'; I love seeing this phenomenon. Let's argue, discuss, and then turn it into serious science with real effort [Quoting @1jaskiratsingh]: ‼️ Representations matter for generation! But it turns out our understanding of how representations help generation has been wrong all along ‼️ Our previous ideas: (we were wrong) ❌ Larger visual encoders → better representations → better generation ❌ Better global semantics → better representations → better generation It turns out: 🤯 For representation alignment, visual encoders more than 20x smaller can achieve similar or better performance than larger models 🤯 Visual encoders with linear probing accuracy of about 20% (a measure of global semantics) can outperform encoders with >80% accuracy 🤯 Even classic features such as SiFT and HoG can bring improvements comparable to much larger modern visual encoders ‼️ 🚨 Introduction: What matters for representation alignment? Global information or spatial structure 🚨 TL;DR: ✅ Better global semantic information ≠ better generation ✅ Spatial structure (not global semantics) drives the generation performance of representations ✅ We propose iREPA: only 3 lines of code, emphasizing spatial structure transfer, and consistently improving convergence speed on methods such as REPA, REPA-E, Meanflow, and JiT An exciting project at @AdobeResearch, with @xingjian_leng, @zongze_wu, @LiangZheng_06, @rzhang88, @elishechtman, and @sainingxie 🙏 For me this was also a particularly fun and unique experience; at every step of the project we were proving our own biases wrong 😆 Also huge thanks to @YouJiacheng, @ShumingHu, and @gallabytes, whose comments on X opened up this direction 🫡 Paper: https://arxiv.org/abs/2512.10794 Code: https://github.com/End2End-Diffusion/iREPA Project page: https://end2end-diffusion.github.io/irepa More details in the thread: [1/n] 🧵

Key Takeaways

  • New paper: iREPA diffusion models are a renderer of their underlying representations.
  • With this new setup,we can gain clearer insight into what these representations really mean.
  • Jas began a spontaneous exploration,and we learned a lot over the past three months.
  • This is also our small experiment with a new kind of online 'water cooler effect'
  • I love seeing this phenomenon.

Quick Facts

Published
Updated
Source
X:谢赛宁 (@sainingxie)
Category
Other
Read the original Read the full article at the original source
Why it matters

颠覆认知:小20倍视觉编码器也能驱动高质量生成,空间结构才是关键

Source & copyright

This article is based on curated titles and AI summaries from the AIHOT public API, with attribution and links back. Original copyright remains with the original publishers and authors; this site does not mirror full texts.

Original article · AIHOT entry

How to Cite

When citing this article, keep the title, date, and permalink, and credit Orange News as the source.

Orange News: new paper: iREPA diffusion models are a renderer of their underlying representations. with this new.... December 16, 2025. https://new.nat6.net/en/news/cmnz6dpf402amsl0fxz3kh21f/

View all →