Research map / Image generation / Fast diffusion sampling
Moving Beyond Diffusion: Hierarchy-to-Hierarchy Autoregression for fMRI-to-Image Reconstruction
Xu Zhang, Ruijie Quan, Wenguan Wang, Yi Yang
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 |
|---|---|---|---|
| supervise every intermediate fMRI encoder block with CLIP-layer MSE (full cascaded supervision) instead of final-block only (ablation) | MindHier · NSD Subject 1 | CLIP score (%): 95.4% -> 97.2% | Improved |
| use hierarchical features (final supervision only) instead of a single terminal feature (ablation) | MindHier · NSD Subject 1 | CLIP score (%): 95.1% -> 95.4% | Improved |
| align with later CLIP layers {18,20,22,24} instead of {12,16,20,24} (ablation) | MindHier · NSD Subject 1 | CLIP score (%): 97.2% -> 95.4% | Hurt |
| align with earlier CLIP layers {6,12,18,24} instead of {12,16,20,24} (ablation) | MindHier · NSD Subject 1 | PixCorr / CLIP: PixCorr 0.273 -> 0.283; CLIP 97.2% -> 94.8% | Mixed |
| invert guidance order (fine-to-coarse) instead of coarse-to-fine scale-aware guidance (ablation) | MindHier · NSD Subject 1 | CLIP score (%): 97.2% -> 96.1% | Hurt |
| use best-of-4 sampling instead of single shot (ablation) | MindHier · NSD | CLIP score (%): 95.0% -> 96.4% | Improved |
| replace diffusion decoding with scale-wise autoregression | MindEye2 pipeline · NSD test | inference time (s): 12.14 -> 2.64 | Improved |
| finetune on one session of fMRI data | MindHier vs MindEye2 · NSD S1 | CLIP score (%): 80.8% (MindEye2) -> 92.2% | Improved |
| cross-subject 1-hour finetuning of MindHier instead of MindEye2 | MindHier · NSD cross-subject 1-hour | CLIP score (%): 80.7% -> 80.2% | Hurt |
| cross-subject 40-hour finetuning of MindHier instead of MindEye2 | MindHier · NSD cross-subject 40-hour | CLIP score (%): 93.5% -> 95.8% | Improved |
| restrict MindEye2 diffusion to 10 denoising steps to match MindHier scales | MindEye2 · NSD qualitative | semantic fidelity | Hurt |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- MindEye2 — MindEye2: Shared-Subject Models Enable fMRI-To-Image With 1 Hour of Data (ICML 2024) · better on 5 of 6 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- High-Resolution Image Reconstruction With Latent Diffusion Models From Human Brain Activity (CVPR 2023 · similarity 0.89)
- Reconstructing the Mind's Eye: fMRI-to-Image with Contrastive Learning and Diffusion Priors (NeurIPS 2023 · similarity 0.88)
- Mind Reader: Reconstructing complex images from brain activities (NeurIPS 2022 · similarity 0.88)
- MindEye2: Shared-Subject Models Enable fMRI-To-Image With 1 Hour of Data (ICML 2024 · similarity 0.88)
- NeuroClips: Towards High-fidelity and Smooth fMRI-to-Video Reconstruction (NeurIPS 2024 · similarity 0.87)
- MoRE-Brain: Routed Mixture of Experts for Interpretable and Generalizable Cross-Subject fMRI Visual Decoding (NeurIPS 2025 · 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.
