Research map / Image generation / Autoregressive and latent generation
Generating metamers of human scene understanding
Ritik Raina, Abe Leite, Alexandros Graikos, Seoyoung Ahn, Dimitris Samaras, Greg Zelinsky
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 |
|---|---|---|---|
| add foveal conditioning to peripheral-only generation (ablation) | MetamerGen · Behavioral same-different task | metamer rate: 45.8% -> 54.5% | Improved |
| add peripheral conditioning to foveal-only generation (ablation) | MetamerGen · Behavioral same-different task | metamer rate: 8.4% -> 54.5% | Improved |
| use viewer's own fixations instead of random fixation locations | MetamerGen · Behavioral same-different task (primary) | metamer rate: 27.7% -> 29.4% | No clear effect |
| increase peripheral blur (greater downsampling) (ablation) | MetamerGen peripheral-only · COCO-10k-test | FID, CLIP similarity, DreamSim: downsampling 0.25x to 112x112 has little effect versus 448x448 | Hurt |
| increase number of foveal tokens (ablation) | MetamerGen foveal-only · COCO-10k-test | CLIP, DreamSim, FID | Improved |
| replace DINOv2 with CLIP as vision encoder (ablation) | MetamerGen peripheral-only · COCO-10k-test | FID | Hurt |
| increase peripheral scale in conditioning (ablation) | MetamerGen · COCO-10k-test (central fixation) | FID | Improved |
| fine-tune on COCO images with DINOv2 conditioning vs text-to-image SD-1.5 | MetamerGen · COCO-10k-test | FID | Improved |
| higher pixel-level similarity (PSNR) between generated and original image | MetamerGen · Behavioral same-different task | same judgment | No clear effect |
Closest papers by meaning
- The Scene Language: Representing Scenes with Programs, Words, and Embeddings (CVPR 2025 · similarity 0.87)
- Compositional Scene Understanding through Inverse Generative Modeling (ICML 2025 · similarity 0.87)
- Diff-2-in-1: Bridging Generation and Dense Perception with Diffusion Models (ICLR 2025 · similarity 0.85)
- Revealing the Gap in Human and VLM Scene Perception through Counterfactual Semantic Saliency (NeurIPS 2026 · similarity 0.85)
- Hummingbird: High Fidelity Image Generation via Multimodal Context Alignment (ICLR 2025 · similarity 0.85)
- Generation Enhances Understanding in Unified Multimodal Models via Multi-Representation Generation (ICML 2026 · similarity 0.85)
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.
