DiT
📌 核心贡献
将 U-Net backbone 替换为 Transformer，提出 Diffusion Transformer（DiT），验证了扩散模型的 scaling law：计算量（GFLOPs）越大，FID 越低。DiT-XL/2 在 ImageNet 256×256 上达到 FID 2.27，是 REPA、iFSQ 等现代生成工作的直接基础。
📖 摘要
We explore a new class of diffusion models based on the transformer architecture. We train latent diffusion models of images, replacing the commonly-used U-Net backbone with a transformer that operates on latent patches. We analyze the scalability of our Diffusion Transformers (DiTs) through the lens of forward pass complexity as measured by Gflops. We find that DiTs with higher Gflops — through increased transformer depth/width or increased number of input tokens — consistently have lower FID. In addition to possessing good scalability properties, our largest DiT-XL/2 models outperform all prior diffusion models on the class-conditional ImageNet 512x512 and 256x256 benchmarks, achieving a state-of-the-art FID of 2.27 on the latter.
🔍 关键信息
| 字段 | 内容 |
|---|---|
| 机构 | UC Berkeley / Meta AI |
| 发表 | 2022-12-19 |
| 分类 | cs.CV, cs.LG |
| 链接 | arXiv |
📝 我的笔记
ICCV 2023. Project page: https://www.wpeebles.com/DiT