Research map / Image generation / Autoregressive and latent generation
Scalable Diffusion Models with Transformers
William Peebles, Saining Xie
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use adaLN-Zero conditioning instead of in-context or cross-attention (ablation) | DiT-XL/2 · ImageNet 256x256, 400K steps | FID | Improved |
| zero-initialize adaLN block (adaLN-Zero) instead of vanilla adaLN (ablation) | DiT-XL/2 · ImageNet 256x256 | FID | Improved |
| decrease patch size (more tokens) holding model size fixed (ablation) | DiT · ImageNet 256x256, 400K steps | FID-50K: DiT-S/8 153.60 -> DiT-S/2 68.40 | Improved |
| increase model size (S to XL) holding patch size fixed (ablation) | DiT · ImageNet 256x256, 400K steps | FID-50K: DiT-S/8 153.60 -> DiT-L/8 118.87 | Improved |
| increase sampling steps to compensate for smaller model compute (ablation) | DiT-L/2 · ImageNet 256x256, 400K steps | FID-10K: XL/2 128 steps 23.7 vs L/2 1000 steps 25.9 | No clear effect |
| add classifier-free guidance (cfg=1.50) | DiT-XL/2 · ImageNet 256x256 | FID-50K: 9.62 -> 2.27 | Improved |
| classifier-free guidance on recall | DiT-XL/2 · ImageNet 256x256 | Recall: 0.67 -> 0.57 | Hurt |
| replace U-Net backbone with transformer on latent patches | latent diffusion · ImageNet 256x256 | FID-50K: 3.60 -> 2.27 | Improved |
Closest papers by meaning
- Dynamic Diffusion Transformer (ICLR 2025 · similarity 0.90)
- Scaling Diffusion Transformers Efficiently via $\mu$P (NeurIPS 2025 · similarity 0.90)
- Exploring Diffusion Transformer Designs via Grafting (NeurIPS 2025 · similarity 0.90)
- Expert Race: A Flexible Routing Strategy for Scaling Diffusion Transformer with Mixture of Experts (ICML 2025 · similarity 0.90)
- Simpler Diffusion: 1.5 FID on ImageNet512 with Pixel-space Diffusion (CVPR 2025 · similarity 0.90)
- Scalable High-Resolution Pixel-Space Image Synthesis with Hourglass Diffusion Transformers (ICML 2024 · similarity 0.89)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
