Research map / Image generation / Autoregressive and latent generation
Diffusion Transformers with Representation Autoencoders
Boyang Zheng, Nanye Ma, Shengbang Tong, Saining Xie
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 |
|---|---|---|---|
| use standard DiT on RAE latents instead of SD-VAE latents (ablation) | DiT-S · ImageNet 256, 80 epochs | gFID: 51.74 -> 215.76 | Hurt |
| widen DiT to at least the RAE token dimension (ablation) | DiT on DINOv2 RAE latents · single-image overfitting | overfitting loss: DiT-S on DINOv2-B 5.2e-1 vs DiT-B on DINOv2-B 2.4e-2 | Improved |
| increase depth from 12 to 24 at width 384 (ablation) | SiT-S on RAE latent · single-image overfitting | sample quality | No clear effect |
| add dimension-dependent noise schedule shift (ablation) | DiT-XL on RAE (DINOv2-B) · ImageNet 256, 80 epochs | gFID: 23.0750453 -> 4.81 | Improved |
| add noise-augmented decoder training (ablation) | RAE decoder · ImageNet 256, 80 epochs | gFID: 4.81 -> 4.28 | Improved |
| add noise-augmented decoder training (ablation) | RAE decoder with MAE-B encoder · ImageNet 256, 80 epochs | gFID: 16.14 -> 8.38 | Improved |
| add shallow wide DDT head to DiT (ablation) | DiT-XL on RAE (DINOv2-B) · ImageNet 256, 80 epochs | gFID: 4.28 -> 2.16 | Improved |
| add shallow wide DDT head to DiT (ablation) | DiT-XL on RAE (DINOv2-L) · ImageNet 256, 80 epochs | gFID: 6.09 -> 2.73 | Improved |
| add DDT head on SD-VAE latents (ablation) | DiT-XL on SD-VAE · ImageNet 256, 80 epochs | gFID: 7.13 -> 11.70 | Hurt |
| diffuse in raw pixels instead of RAE latents (ablation) | DiT-DH-XL · ImageNet 256, 80 epochs | gFID: 2.16 -> 30.56 | Hurt |
| use decoder upsampling from 256 tokens instead of direct 512 resolution training (ablation) | RAE DiT-DH-XL · ImageNet 512, 400 epochs | gFID: 1.13 -> 1.61 | Mixed |
| use larger decoder (ViT-B to ViT-XL) (ablation) | RAE DINOv2-B · ImageNet 256 reconstruction | rFID: 0.58 -> 0.49 | Improved |
Closest papers by meaning
- Reconstruction vs. Generation: Taming Optimization Dilemma in Latent Diffusion Models (CVPR 2025 · similarity 0.92)
- Representation Alignment for Generation: Training Diffusion Transformers Is Easier Than You Think (ICLR 2025 · similarity 0.91)
- Improved Baselines with Representation Autoencoders (NeurIPS 2026 · similarity 0.90)
- Representation Entanglement for Generation: Training Diffusion Transformers Is Much Easier Than You Think (NeurIPS 2025 · similarity 0.90)
- Boosting Generative Image Modeling via Joint Image-Feature Synthesis (NeurIPS 2025 · similarity 0.89)
- Masked Autoencoders Are Effective Tokenizers for Diffusion Models (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.
