Research map / Image generation / Autoregressive and latent generation
$\textit{MADFormer}$: Mixed Autoregressive and Diffusion Transformers for Continuous Image Generation
Junhao Chen, Yulia Tsvetkov, Xiaochuang Han
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 |
|---|---|---|---|
| allocate more layers to AR (e.g. 3:1 AR:diffusion) under low inference NFE (ablation) | MADFormer · FFHQ-1024 and ImageNet, low NFE | FID: 60-75% lower FID | Improved |
| allocate more layers to diffusion when compute is high (ablation) | MADFormer · high NFE, high resolution | FID | Improved |
| increase diffusion depth from 7 to 28 layers (ablation) | MADFormer · FFHQ-1024 | FID: 20.2 -> 15.9 | Improved |
| increase diffusion depth from 7 to 28 layers (ablation) | MADFormer · ImageNet 256 | FID: 34.0 -> 27.4 | Improved |
| split FFHQ-1024 into 16 blocks instead of 4 or 64 (ablation) | MADFormer · FFHQ-1024 | FID: 18.9 -> 17.8 | Improved |
| increase AR length (more blocks) (ablation) | MADFormer · ImageNet 256 | FID: 28.4 -> 39.7 | Hurt |
| add AR condition on top of clean blocks (ablation) | MADFormer · FFHQ-1024 | FID: 20.1 -> 17.8 | Improved |
| add clean blocks on top of AR condition (ablation) | MADFormer · ImageNet 256 | FID: 31.2 -> 30.0 | Improved |
| use separate parameter sets for text, clean and noisy blocks (ablation) | MADFormer · FFHQ-1024 and ImageNet | FID: ImageNet 30.4 -> 30.0 | No clear effect |
| keep sequence-level causal attention during diffusion instead of within-block MLP-style denoising (ablation) | MADFormer · ImageNet 256 | FID: 96.5 -> 30.0 | Improved |
| add hidden loss (lambda 0.1) (ablation) | MADFormer · FFHQ-1024 | FID: 19.4 -> 17.8 | Improved |
| raise hidden loss weight from 0.1 to 1 (ablation) | MADFormer · FFHQ-1024 | FID: 17.8 -> 18.4 | Hurt |
Closest papers by meaning
- Denoising Autoregressive Transformers for Scalable Text-to-Image Generation (ICLR 2025 · similarity 0.91)
- Autoregressive Distillation of Diffusion Transformers (CVPR 2025 · similarity 0.90)
- HART: Efficient Visual Generation with Hybrid Autoregressive Transformer (ICLR 2025 · similarity 0.90)
- Revisiting Non-Autoregressive Transformers for Efficient Image Synthesis (CVPR 2024 · similarity 0.90)
- ControlAR: Controllable Image Generation with Autoregressive Models (ICLR 2025 · similarity 0.90)
- Visual Autoregressive Modeling: Scalable Image Generation via Next-Scale Prediction (NeurIPS 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.
