Research map / Image generation / Autoregressive and latent generation
Compositional Visual Generation with Energy Based Models
Yilun Du, Shuang Li, Igor Mordatch
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 more energy functions via conjunction (Pos & Color vs single concepts) | EBM · MuJoCo Scenes | position/color accuracy: Pos acc 0.984 -> 0.801; Color acc 0.997 -> 0.8125 | Mixed |
| negate color concept in conjunction with position (ablation) | EBM · MuJoCo Scenes | color accuracy: 0.997 -> 0.096 | Hurt |
| negate position concept in conjunction with color (ablation) | EBM · MuJoCo Scenes | position accuracy: 0.984 -> 0.033 | Hurt |
| compose independently trained EBMs vs product-of-latents approach [29] | conjunction of position and color · MuJoCo Scenes | position accuracy: 0.151 -> 0.801 | Improved |
| composed EBMs vs SNGAN trained on all four attributes | CelebA generation · CelebA 4 attributes | FID: 64.5 -> 45.3 | Improved |
| compose fixed earlier EBMs with new EBMs instead of sequentially training a GAN | continual learning · MuJoCo continual learning | position accuracy after learning color: 0.117 -> 0.781 | Improved |
| EBM vs GAN on newest concept (color) after continual learning | continual learning · MuJoCo continual learning | color accuracy: 0.984 -> 0.521 | Hurt |
| factorized EBM composition vs holistic baseline in cross-product extrapolation | sphere size/position generation · MuJoCo extrapolation | size and position error | Improved |
| increase number of images in conjunction for concept inference | positional EBM · MuJoCo position inference | position error | Improved |
Closest papers by meaning
- ConceptMix: A Compositional Image Generation Benchmark with Controllable Difficulty (NeurIPS 2024 · similarity 0.86)
- Compositional Scene Understanding through Inverse Generative Modeling (ICML 2025 · similarity 0.86)
- Does Data Scaling Lead to Visual Compositional Generalization? (ICML 2025 · similarity 0.86)
- Iterated Learning Improves Compositionality in Large Vision-Language Models (CVPR 2024 · similarity 0.85)
- ConceptBed: Evaluating Concept Learning Abilities of Text-to-Image Diffusion Models (AAAI 2024 · similarity 0.85)
- FreeCustom: Tuning-Free Customized Image Generation for Multi-Concept Composition (CVPR 2024 · 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.
