arXiv cs.AI· Denis Lukovnikov, Andreas M\"uller, Asja Fischer·· 10 小时前AI 评分31
JiT 图像 Transformer 引入后期像素精修，ImageNet@256 FID 降至 1.78
Just Image Transformers with Late-Stage Pixel Refinement
AI 导读
研究者在 JiT 的 post-training 阶段加入简单卷积像素通路层，显著提升无 latent VAE 的图像生成质量。该像素通路无需在去噪早期启用，可延迟到最后 20% 的去噪过程才激活。ImageNet@256 实验显示 FID 从修正采样后的 2.34 降至 1.78，且未使用额外损失。
正文 · 原文
Abstract:Recent work on latent-free diffusion transformers has demonstrated the feasibility of training image generators without a latent VAE. This avoids compression artifacts from the VAE and simplifies the training pipeline. In particular, JiT showed that even a standard transformer can be trained to generate images when the model is trained to predict the final image ($x$-prediction). Even though JiT unlocks training efficiency comparable to latent models, it still suffers from reduced image quality likely arising from modeling the image as a collection of large patches. In this work, we revisit JiT and show that adding simple convolutional pixel-pathway layers in a post-training phase significantly improves generation quality. Moreover, we show that the added pixel-pathway does not need to be active in the early denoising stages, and can be activated as late as in the last $20\%$ of the denoising process. Our experiments on ImageNet@256 show improvements compared to the original JiT results, achieving an FID of $1.78$ (from $2.34$ with corrected sampling), without using additional losses.
| Subjects: | Computer Vision and Pattern Recognition (cs.CV); Artificial Intelligence (cs.AI) |
| Cite as: | arXiv:2611.13120 [cs.CV] |
| (or arXiv:2611.13120v1 [cs.CV] for this version) | |
| https://doi.org/10.48550/arXiv.2611.13120 arXiv-issued DOI via DataCite (pending registration) |
Submission history
From: Denis Lukovnikov [view email]
[v1]
Mon, 9 Nov 2026 19:08:13 UTC (37,621 KB)
来源：arXiv cs.AI · arxiv.org