Summary
This post introduces a computer vision paper by Sihan Xu, Ji Xie, Zilin Wang, Hui Shen, and Stella X. Yu, available on arXiv as 2610.02203. In diffusion transformers, a class label or text prompt is typically embedded once and the same condition is reused at every denoising step. The proposed method, Next-Embedding Predictive Autoregression (NEPA), instead trains a Transformer to predict the next continuous embedding in a sequence. During generation, the clean image follows the noisy image, so its embeddings are the next embeddings after the condition and the noisy image. A NEPA model predicts all of these embeddings at once via Multi-Embedding Prediction, and a DiT generator is conditioned on these predictions in Embedding Conditioned Generation, with conditions recomputed at every denoising step so the signal adapts to the current noisy state. Experiments on class-conditional ImageNet 256x256 study conditioning of the generator, the design of multi-embedding prediction, and scaling of both models. NEPA adds a second network per sampling step; combined with REPA, the final NEPA-DiT-XL model reaches an FID of 1.32 using roughly one-third of REPA's training compute.
Paper Overview
- Research area: Computer Vision
- Authors: Sihan Xu, Ji Xie, Zilin Wang, Hui Shen, Stella X. Yu
- Published: 2026-10-01
- arXiv: 2610.02203
Summary
In diffusion transformers, a class label or a text prompt is embedded once, and the same condition is reused at every denoising step. The authors ask whether predicted embeddings can serve as this condition instead.
Next-Embedding Predictive Autoregression (NEPA) trains a Transformer to predict the next continuous embedding in a sequence. In generation, the clean image follows the noisy image, so its embeddings are the next embeddings after the condition and the noisy image.
Key components:
- Multi-Embedding Prediction: a NEPA model is trained to predict all of the clean image's embeddings at once.
- Embedding Conditioned Generation: a DiT generator is conditioned on these predictions, recomputed at every denoising step, so the conditioning signal adapts to the current noisy state.
Experiments and Results
Experiments on class-conditional ImageNet \(256\times256\) generation study the conditioning of the generator, the design of multi-embedding prediction, and the scaling of both models. NEPA adds a second network for each sampling step. Combined with REPA, the final model NEPA-DiT-XL achieves an FID of 1.32 with roughly one-third of REPA's training compute.
---
*Automatically collected on 2026-10-03.*
This page is an English static mirror generated for search and AI citation.
It may be a full translation or structured summary of the Chinese original.
Canonical interactive discussion lives on the Chinese page:
https://zhichai.net/topic/178635510