Research map / Image generation / Fast diffusion sampling
Local Mechanisms of Compositional Generalization in Conditional Diffusion
Arwen Bradley
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 |
|---|---|---|---|
| enforce local conditional scores with a patch-local architecture (Exp.2L vs Exp.2) (ablation) | single-object grid-conditioned CLEVR model, trained on 1-3 objects · Location-conditioned CLEVR | K_max: 2 -> 9 | Improved |
| enforce local architecture (Exp.2L vs Exp.2), trained on 1 object (ablation) | single-object grid-conditioned model · Location-conditioned CLEVR, trained on 1 object | K_max: 1 -> 5 | Improved |
| enforce local architecture (Exp.2L vs Exp.2), trained on 1-5 objects (ablation) | single-object grid-conditioned model · Location-conditioned CLEVR, trained on 1-5 objects | K_max: 3 -> 10 | Improved |
| enforce local architecture (Exp.3L vs Exp.3), trained on 1-3 objects (ablation) | list-conditioned model · Location-conditioned CLEVR, trained on 1-3 objects | K_max: 3 -> 9 | Improved |
| label all object locations with grid conditioner instead of a single object (Exp.1 vs Exp.2), 1-3 objects | grid-conditioned CLEVR model · Location-conditioned CLEVR, trained on 1-3 objects | K_max: 2 -> 7 | Improved |
| use grid-style instead of list-style conditioner (Exp.1 vs Exp.3), 1-3 objects | location-conditioned CLEVR model · Location-conditioned CLEVR, trained on 1-3 objects | K_max: 3 -> 7 | Improved |
| train on more objects (1 to 1-3 to 1-5) for Exp.1 | grid-conditioned model · Location-conditioned CLEVR | K_max: 1 -> 7 -> 10 | Improved |
| train on more objects (1-3 to 1-5) for the list-conditioned model | Exp.3 · Location-conditioned CLEVR | K_max: 3 -> 6 | Improved |
| feature-space instead of pixel-space for disentanglement of concepts | SDXL · SDXL hidden activations, 24 prompts | disentanglement ratio: 1.42 -> 1.4--1.6 | Improved |
Closest papers by meaning
- How Diffusion Models Learn to Factorize and Compose (NeurIPS 2024 · similarity 0.90)
- How Compositional Generalization and Creativity Improve as Diffusion Models are Trained (ICML 2025 · similarity 0.90)
- When does compositional structure yield compositional generalization? A kernel theory. (ICLR 2025 · similarity 0.89)
- Does Data Scaling Lead to Visual Compositional Generalization? (ICML 2025 · similarity 0.89)
- Diffusion Classifiers Understand Compositionality, but Conditions Apply (NeurIPS 2025 · similarity 0.88)
- Swing-by Dynamics in Concept Learning and Compositional Generalization (ICLR 2025 · similarity 0.88)
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.
