Research map / Image generation / Autoregressive and latent generation
Autoregressive Image Diffusion: Generation of Image Sequence and Application in MRI
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes autoregressive image diffusion (AID), a diffusion model over image sequences with causal attention over previous images, and uses it as a posterior-sampling prior for accelerated MRI reconstruction on fastMRI, outperforming a standard diffusion model (Guide) and CSGM/NCSNv2.
Problem
Accelerated MRI reconstruction from undersampled k-space suffers from artifacts and hallucinations; standard diffusion priors treat images independently and ignore inter-image dependencies in clinical image sequences/volumes.
Approach
Combine an autoregressive factorization over image sequences with DDPM-style diffusion: each image's noise prediction is conditioned on previous clean images via causal attention in a U-Net (temporal-spatial conditioning), trained in parallel, and used as a prior for DDIM-based posterior sampling with a k-space data-fidelity step.
Key results
- On 12x undersampled fastMRI validation volumes, AID outperforms Guide and NCSNv2 in PSNR and NRMSE, with the largest gain without ACS.
- AID reduces folding artifacts in single-coil unfolding compared with the Guide model and produces fewer visible hallucinations in reconstructions.
- AID generates sequentially coherent brain, cardiac, and UAV image sequences, including from a cold start (black initial background).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use autoregressive image diffusion instead of standard diffusion as the MRI reconstruction prior | MRI reconstruction from 12-times undersampled fastMRI k-space · fastMRI validation, random and equispaced masks with and without ACS | PSNR | Helped |
| Use autoregressive image diffusion instead of standard diffusion as the MRI reconstruction prior | MRI reconstruction from 12-times undersampled fastMRI k-space · fastMRI validation, random and equispaced masks with and without ACS | NRMSE | Helped |
| Learn inter-image dependencies with AID rather than use Guide | MRI reconstruction · Single-coil unfolding with two-times undersampling | reconstruction error in folding region | Helped |
| Learn inter-image dependencies with AID rather than use Guide | MRI reconstruction · Undersampled k-space reconstruction | hallucination severity | Helped |
| Generate images autoregressively conditioned on previous images | AID · Image sequence generation | sequential coherence | Helped |
| Initialize the prospective sampling window with zeros rather than a provided sequence | AID · Prospective cold-start sampling | sequential coherence | Mixed |
| Use two-stage training: train the U-net, then train the temporal-spatial conditioning block with the U-net frozen | AID on ABIDE · ABIDE | training efficiency | Helped |
Limitations
- No evaluation on common natural image benchmarks (ImageNet, CIFAR-10) and no FID or Inception Score reported.
- Evaluation of MRI reconstruction limited to 20 validation volumes at one undersampling factor (12x) in the main comparison.
- Natural-image UAV samples show striped artifacts in temporal x-t planes.
Experimental setup
Datasets: fastMRI brain (T1, T2, FLAIR) cardiac cine (SSA-FARY reconstructions) ABIDE UAV view dataset
Benchmarks: fastMRI validation (20 volumes, 12x undersampling)
Metrics: PSNR NRMSE
Baselines: Guide (standard diffusion model) CSGM NCSNv2
Builds on: DDPM DDIM Score-based generative models (CSGM) DiT guided diffusion
Closest papers by meaning
- Ambient Diffusion Posterior Sampling: Solving Inverse Problems with Diffusion Models Trained on Corrupted Data (ICLR 2025 · similarity 0.88)
- Improving Diffusion Posterior Samplers with Lagged Temporal Corrections for Image Restoration (NeurIPS 2026 · similarity 0.87)
- Rethinking Diffusion Model for Multi-Contrast MRI Super-Resolution (CVPR 2024 · similarity 0.87)
- Improving Diffusion-Based Image Restoration with Error Contraction and Error Correction (AAAI 2024 · similarity 0.87)
- Autoregressive Distillation of Diffusion Transformers (CVPR 2025 · similarity 0.87)
- Boosting Diffusion Models with Moving Average Sampling in Frequency Domain (CVPR 2024 · similarity 0.87)
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.