Scalable Diffusion Models with TransformersScalable Diffusion Models with Transformers
问题:扩散模型能否用 Transformer 替代 U-Net 作为骨干,并像 Gflops 一样可扩展。
证据位置:摘要在潜空间扩散中用 Transformer 处理 latent patches;Gflops 更高的 DiT(更深/更宽或更多 token)FID 更低;DiT-XL/2 在 ImageNet 256 上 FID 2.27。
作者自陈局限:ImageNet 类条件生成;FID 单一指标;训练成本高;文本/视频等通用生成未验证。
它没主张的:不是通用生成模型;不是 FID 唯一标准;不是替代所有 U-Net。Problem: can diffusion models use a Transformer instead of a U-Net backbone and scale like Gflops?
Where the evidence sits: the abstract replaces the U-Net with a transformer over latent patches in latent diffusion; higher-Gflops DiTs (deeper/wider or more tokens) consistently have lower FID; DiT-XL/2 reaches FID 2.27 on ImageNet 256.
What the authors say about limits: class-conditional ImageNet generation; FID is a single metric; training is expensive; text/video and other general generation are not validated.
What it does not claim: not a general generative model; not FID as the only criterion; not a replacement for every U-Net.
关键参数Key specs
- 证据位置 / evidence
- 摘要:latent patches、Gflops、FID 2.27
- FID / ImageNet 256
- 2.27(DiT-XL/2)
- 复现状态 / reproduction
- 未复现(只读摘要)
官方原文怎么说What the source says
replacing the commonly-used U-Net backbone with a transformer that operates on latent patches
DiTs with higher Gflops ... consistently have lower FID
achieving a state-of-the-art FID of 2.27 on the latter
资料与许可Source and licence
- 资料时间Source checked
- 2026-11-19
- 许可 / 类型Licence / kind
- arxiv-preprint
- 许可说明Licence evidence
- https://arxiv.org/abs/2212.09748
- 类目Category
- 架构Architecture
论文里的数字怎么读How to read the numbers
先分清测什么、怎么算、和谁比、有没有数据污染。拿不准就回原文看实验设置。Before citing a score, separate what it measures, how it is scored, what it compares against and whether contamination is controlled. If unsure, go back to the experiment setup in the paper. 看评测基准See benchmark cards · 看数据集卡See dataset cards