Research map / Image generation / Fast diffusion sampling
D-AR: Diffusion via Autoregressive Models
Ziteng Gao, Mike Zheng Shou
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 linear CFG schedule instead of no CFG (ablation) | D-AR-L · ImageNet 256x256 class-conditional | gFID: 7.43 -> 2.44 | Improved |
| use linear CFG schedule instead of constant 1.5 (ablation) | D-AR-XL · ImageNet 256x256 class-conditional | gFID: 3.39 -> 2.09 | Improved |
| increase time-shift beta from 1 to 2 (ablation) | sequential diffusion tokenizer · ImageNet val (lightweight tokenizer) | rFID: 39.75 -> 28.65 | Improved |
| increase time-shift beta from 2 to 4 (ablation) | sequential diffusion tokenizer · ImageNet val (lightweight tokenizer) | rFID: 28.65 -> 27.10 | No clear effect |
| increase conditioning groups K from 8 to 16 (ablation) | sequential diffusion tokenizer · ImageNet val (lightweight tokenizer) | rFID: 28.65 -> 49.07 | Hurt |
| decrease conditioning groups K from 8 to 4 (ablation) | sequential diffusion tokenizer · ImageNet val (lightweight tokenizer) | rFID: 28.65 -> 23.04 | Improved |
| use K=8 vs K=4 tokens for AR training (ablation) | D-AR-B · ImageNet 256x256 | gFID: 53.70 -> 54.26 | No clear effect |
| use K=8 vs K=16 tokens for AR training (ablation) | D-AR-B · ImageNet 256x256 | gFID: 63.69 -> 54.26 | Improved |
| generate more AR tokens (64 to 256) before decoding | D-AR-L · ImageNet 256x256 | gFID: 7.38 -> 2.44 | Improved |
| replace LlamaGen tokenizer with sequential diffusion tokenizer | tokenizer, 256 tokens · ImageNet val | rFID | Improved |
| do not use adversarial loss (ablation) | sequential diffusion tokenizer · ImageNet | training stability | No clear effect |
Closest papers by meaning
- Autoregressive Distillation of Diffusion Transformers (CVPR 2025 · similarity 0.91)
- Visual Autoregressive Modeling: Scalable Image Generation via Next-Scale Prediction (NeurIPS 2024 · similarity 0.91)
- Denoising Autoregressive Transformers for Scalable Text-to-Image Generation (ICLR 2025 · similarity 0.91)
- ControlAR: Controllable Image Generation with Autoregressive Models (ICLR 2025 · similarity 0.90)
- Faster Diffusion: Rethinking the Role of the Encoder for Diffusion Model Inference (NeurIPS 2024 · similarity 0.90)
- BitDance: Scaling Autoregressive Generative Models with Binary Tokens (NeurIPS 2026 · similarity 0.90)
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.
