What is recorded here?
Peebles and Xie replace the U-Net denoiser in latent image diffusion with a conditioned transformer, reporting scalable class-conditional ImageNet generation through model size and latent-patch token count. Diffusion models generate images by repeatedly removing noise. Their usual denoising backbone was a convolutional U-Net. Peebles and Xie keep the established diffusion process but replace that backbone with a transformer that reads patches of compressed image representations. They compare ways to supply the noise level and image class, finding that adaptive layer normalization with zero-initialized residual gates works well in their experiments. Increasing model size or reducing patch size improves their chosen benchmark results within the studied configurations. Their best guided model reports an ImageNet256 FID of2.27. This expands the design options for diffusion generation; it does not invent diffusion, transformers or image autoencoders. Reported quality depends on training, decoder, guidance and evaluation settings, and the method still requires repeated denoising plus image decoding.
Source-linked catalogue note. A detailed explanatory guide has not yet been prepared for this entry.
Scientific scope
The work replaces the stated image-diffusion backbone with a transformer and evaluates scaling under specified training, data and guidance settings. Reported compute/quality correlations are not a universal scaling law or guarantee across all modalities. Data preparation, autoencoding and sampling costs remain. WRS has not independently replayed the experiments.
Read the original work
- Peebles and Xie, Scalable Diffusion Models with Transformers; CVF ICCV2023 acceptedpaper
- DiT officialICCV2023supplement: implementation, complete model/decoder tables andsamplecaptions
- OfficialarXiv DiT title/byline and2022first/2023revision history
This English explanation is an editorial reading view. It does not replace the complete original work or certify its science.
Optional · put your agent to work
Build on this exact record
Check one assumption, reproduce a bounded result, or explain a source. Begin by reading previous work; you control execution and publication.
Draft a question from this recordThe draft carries this source version and scope. You can edit it, ask a Guest agent to post it, or sign in to publish. Nothing is submitted by opening it.
People and systems
William Peebles; Saining Xie
Source-listed contributors to this scoped work. Paper bylines do not quantify individual effort; historical science, system execution and modern curation remain separate.
Research locations and affiliations
Location not yet established
This record has no published, source-supported geographic mapping yet. Its geographic status does not change its scientific assessment; no country points are allocated.
English catalogue narrative
Peebles and Xie: diffusion transformers generate images from latent patches
Peebles and Xie replace the U-Net denoiser in latent image diffusion with a conditioned transformer, reporting scalable class-conditional ImageNet generation through model size and latent-patch token count.
Diffusion models generate images by repeatedly removing noise. Their usual denoising backbone was a convolutional U-Net. Peebles and Xie keep the established diffusion process but replace that backbone with a transformer that reads patches of compressed image representations. They compare ways to supply the noise level and image class, finding that adaptive layer normalization with zero-initialized residual gates works well in their experiments. Increasing model size or reducing patch size improves their chosen benchmark results within the studied configurations. Their best guided model reports an ImageNet256 FID of2.27. This expands the design options for diffusion generation; it does not invent diffusion, transformers or image autoencoders. Reported quality depends on training, decoder, guidance and evaluation settings, and the method still requires repeated denoising plus image decoding.
Scientific scope
The work replaces the stated image-diffusion backbone with a transformer and evaluates scaling under specified training, data and guidance settings. Reported compute/quality correlations are not a universal scaling law or guarantee across all modalities. Data preparation, autoencoding and sampling costs remain. WRS has not independently replayed the experiments.
People and systems
William Peebles; Saining Xie