Research map / Image generation / Autoregressive and latent generation
All Are Worth Words: A ViT Backbone for Diffusion Models
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 U-ViT-S/2 instead of GenViT | ViT diffusion backbone · CIFAR10 | FID: 20.20 -> 3.11 | Helped |
| Increase model capacity from U-ViT-M/4 to U-ViT-L/4 | U-ViT · ImageNet 64×64 | FID: 5.85 -> 4.26 | Helped |
| Use latent diffusion with U-ViT-H/2 rather than LDM | ImageNet 256×256 class-conditional generation · ImageNet 256×256 | FID: 3.60 -> 2.29 | Helped |
| Increase depth from 13 to 17 layers | U-ViT-S/2 · MS-COCO validation 256×256 | FID: 5.95 -> 5.48 | Helped |
| Replace the U-Net backbone with U-ViT | Comparable-size latent diffusion model · MS-COCO validation 256×256 | FID: 7.32 -> 5.95 | Helped |
| Replace the parameter- and compute-comparable U-Net with U-ViT | ImageNet 256×256 diffusion model without classifier-free guidance · ImageNet 256×256 | FID: 10.69 -> 6.58 | Helped |
| Replace the parameter- and compute-comparable U-Net with U-ViT | ImageNet 256×256 diffusion model with guidance scale 0.4 · ImageNet 256×256 | FID: 2.66 -> 2.29 | Helped |
| Add a long skip branch using concatenation followed by linear projection rather than omitting the long skip (ablation) | U-ViT · CIFAR10 | FID | Helped |
| Add the long skip by direct addition instead of omitting it (ablation) | U-ViT · CIFAR10 | FID | No clear effect |
| Feed time as a token rather than through adaptive layer normalization (ablation) | U-ViT · CIFAR10 | FID | Helped |
| Add a 3×3 convolution after projection to image patches rather than before projection or omit it (ablation) | U-ViT · CIFAR10 | FID | Helped |
| Use linear patch embedding instead of convolutional patch embedding (ablation) | U-ViT · CIFAR10 | FID | Helped |
Closest papers by meaning
- UDT: Reconciling U-Nets and Diffusion Transformers with Data-Adaptive Token Reduction (NeurIPS 2026 · similarity 0.90)
- Reconstruction vs. Generation: Taming Optimization Dilemma in Latent Diffusion Models (CVPR 2025 · similarity 0.89)
- LaVin-DiT: Large Vision Diffusion Transformer (CVPR 2025 · similarity 0.89)
- DiCo: Revitalizing ConvNets for Scalable and Efficient Diffusion Modeling (NeurIPS 2025 · similarity 0.89)
- U-DiTs: Downsample Tokens in U-Shaped Diffusion Transformers (NeurIPS 2024 · similarity 0.89)
- ViDiT-Q: Efficient and Accurate Quantization of Diffusion Transformers for Image and Video Generation (ICLR 2025 · similarity 0.88)
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.