Research map / Image generation / Fast diffusion sampling
Autoregressive 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 an order-agnostic autoregressive diffusion process instead of absorbing D3PM | text8 character modelling · text8 | NLL (bpc): 1.45 -> 1.43 | Helped |
| Use a 20-step parallel schedule rather than 20-step absorbing D3PM | text8 generation · text8 | NLL (bpc): 1.56 ±0.040 -> 1.51 ±0.007 | Helped |
| Use depth upscaling with factor 4 instead of order-agnostic generation without upscaling (ablation) | ARDM on CIFAR-10 · CIFAR-10 | NLL (bpd): 2.69 ± 0.005 -> 2.64 ± 0.002 | Helped |
| Parallelize generation to a 50-step schedule (ablation) | ARDM-OA on CIFAR-10 · CIFAR-10 | NLL (bpd): 2.69 ± 0.005 -> 2.74 | Hurt |
| Parallelize generation to a 4 × 50-step schedule (ablation) | ARDM-Upscale 4 on CIFAR-10 · CIFAR-10 | NLL (bpd): 2.64 ± 0.002 -> 2.68 | Hurt |
| Use upscale factor 4 rather than factor 16 or 256 (ablation) | ARDM on SC09 · SC09 test | Performance (bpd): 6.36 -> 6.29 | Helped |
| Use upscale factor 2 instead of factor 4 (ablation) | ARDM on SC09 · SC09 test | Performance (bpd): 6.29 -> 6.29 | No clear effect |
| Use upscale factor 4 rather than factor 2 or 16 (ablation) | ARDM on CIFAR-10 · CIFAR-10 | Performance (bpd): 2.67 -> 2.64 | Helped |
| Add depth upscaling with factor 4 (ablation) | ARDM on SC09 · SC09 test | Performance (bpd): 7.93 -> 6.29 | Helped |
| Reduce the SC09 generation budget from 8 × 16000 to 8 × 100 steps (ablation) | ARDM Upscale 4 · SC09 test | Performance (bpd): 6.29 -> 6.32 | Hurt |
| Use ARDM-Upscale 4 rather than IDF++ for per-image coding | CIFAR-10 lossless compression · CIFAR-10 compression per image | Compression (bpd): 3.26 -> 2.71 | Helped |
| Use ARDM-Upscale 4 rather than VDM for dataset coding | CIFAR-10 lossless compression · CIFAR-10 dataset compression | Compression (bpd): 2.72 -> 2.71 | No clear effect |
Closest papers by meaning
- Diffusion Beats Autoregressive in Data-Constrained Settings (NeurIPS 2025 · similarity 0.89)
- Autoregressive Distillation of Diffusion Transformers (CVPR 2025 · similarity 0.88)
- Block Diffusion: Interpolating Between Autoregressive and Diffusion Language Models (ICLR 2025 · similarity 0.87)
- Retrieval-Augmented Diffusion Models for Time Series Forecasting (NeurIPS 2024 · similarity 0.86)
- Neural Diffusion Models (ICML 2024 · similarity 0.86)
- Simple and Effective Masked Diffusion Language Models (NeurIPS 2024 · similarity 0.86)
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.