Research map / Image generation / Autoregressive and latent generation
DiffiT: Diffusion Vision Transformers for Image Generation
Ali Hatamizadeh, Jiaming Song, Guilin Liu, Jan Kautz, Arash Vahdat
Read the paperCheck your idea against 137,217 papers
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 |
|---|---|---|---|
| add time-dependent multi-head self-attention (TMSA) in place of standard self-attention in DDPM++ (VE) (ablation) | DDPM++ (VE) · CIFAR-10 | FID: 3.77 -> 3.49 | Improved |
| add time-dependent multi-head self-attention in DDPM++ (VP) (ablation) | DDPM++ (VP) · CIFAR-10 | FID: 3.01 -> 2.76 | Improved |
| add DiffiT transformer blocks to encoder (ablation) | multi-resolution U-shaped ViT · CIFAR-10 | FID: 3.71 -> 2.27 | Improved |
| add DiffiT transformer blocks to decoder (ablation) | multi-resolution U-shaped ViT · CIFAR-10 | FID: 2.27 -> 1.95 | Improved |
| use multi-resolution (downsampling/upsampling) stages instead of isotropic ViT (ablation) | ViT encoder and SETR-MLA decoder · CIFAR-10 | FID: 5.34 -> 3.71 | Improved |
| inject time token into q,k,v projections rather than only relative position bias (ablation) | DiffiT · CIFAR-10 | FID: 3.97 -> 1.95 | Improved |
| inject time token into attention rather than only the MLP layer (ablation) | DiffiT · CIFAR-10 | FID: 3.81 -> 1.95 | Improved |
| mix time and spatial tokens in the same attention rather than attending separately (ablation) | DiffiT · CIFAR-10 | FID: 2.28 -> 1.95 | Improved |
| use positional instead of Fourier time embedding (ablation) | DiffiT · CIFAR-10 | FID: 2.02 -> 1.95 | Improved |
| increase classifier-free guidance scale beyond 4.6 (256) / 1.49 (512) (ablation) | latent DiffiT · ImageNet-256 / ImageNet-512 | FID: best 1.73 (256) and 2.67 (512) | Hurt |
| use power-cosine improved classifier-free guidance schedule (ablation) | latent DiffiT · ImageNet-512 | FID | No clear effect |
| use DiffiT instead of DiT-XL/2-G backbone design | latent diffusion on ImageNet-256 · ImageNet-256 | parameters (M) / FID: 675 -> 561 params; FID 2.27 -> 1.73 | Improved |
Later papers that compared with this method
7 later accepted papers report DiffiT as a baseline in their results tables; 4 of them report a better number on most of the results they share with it.
- FGM-HD: Boosting Generation Diversity of Fractal Generative Models through Hausdorff Dimension Induction (AAAI 2026 · better on 1 of 2 shared results)
- Condition Errors Refinement in Autoregressive Image Generation with Diffusion Loss (ICLR 2026 · better on 1 of 1 shared result)
- Frequency-Aware Flow Matching for High-Quality Image Generation (CVPR 2026 · better on 1 of 1 shared result)
- CART: Compositional AutoRegressive Transformer for Image Generation (AAAI 2026 · better on 1 of 1 shared result)
- DyDiT++: Diffusion Transformers With Timestep and Spatial Dynamics for Efficient Visual Generation (TPAMI 2026 · better on 0 of 1 shared result)
- Alleviating Distortion in Image Generation via Multi-Resolution Diffusion Models and Time-Dependent Layer Normalization (NeurIPS 2024 · better on 1 of 2 shared results)
- Autoregressive Image Generation without Vector Quantization (NeurIPS 2024 · better on 2 of 2 shared results)
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- SD-DiT: Unleashing the Power of Self-supervised Discrimination in Diffusion Transformer (CVPR 2024 · similarity 0.91)
- DiCo: Revitalizing ConvNets for Scalable and Efficient Diffusion Modeling (NeurIPS 2025 · similarity 0.91)
- LaVin-DiT: Large Vision Diffusion Transformer (CVPR 2025 · similarity 0.89)
- DiT4Edit: Diffusion Transformer for Image Editing (AAAI 2025 · similarity 0.89)
- Layer- and Timestep-Adaptive Differentiable Token Compression Ratios for Efficient Diffusion Transformers (CVPR 2025 · similarity 0.89)
- DCTdiff: Intriguing Properties of Image Generative Modeling in the DCT Space (ICML 2025 · 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.
