Research map / Image generation / Fast diffusion sampling
Compositional Abilities Emerge Multiplicatively: Exploring Diffusion Models on a Synthetic Task
Maya Okawa, Ekdeep Singh Lubana, Robert P. Dick, Hidenori Tanaka
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 |
|---|---|---|---|
| Increase concept distance of test class from training classes | Conditional diffusion model · Synthetic shapes | learning speed of compositional generalization | Hurt |
| Measure compositional accuracy multiplicatively instead of additively | Class 111 accuracy · Synthetic shapes | accuracy dynamics | Mixed |
| Add a fourth concept variable (background color) | Conditional diffusion model · Synthetic shapes, 4 concept variables | accuracy by concept distance | No clear effect |
| Increase samples of concept class 001 in training data from 0 to 300 (ablation) | Conditional diffusion model · Synthetic shapes, class 001 | accuracy of color capability: approx. 10 samples | Improved |
| Increase samples of concept class 100 in training data from 0 to 300 (ablation) | Conditional diffusion model · Synthetic shapes, class 100 | accuracy of shape capability | Improved |
| Fine-tune on added nodes 101, 010, 011 after adversarial training (ablation) | Conditional diffusion model · Adversarial task | generation of correct small circles | Hurt |
| Train on non-neighboring concept classes (adversarial setup) (ablation) | Conditional diffusion model · Adversarial task | compositional generalization | Hurt |
| Underrepresented gender class in CelebA training data | Conditional diffusion model · CelebA | convergence speed of gender accuracy | Hurt |
Closest papers by meaning
- How Compositional Generalization and Creativity Improve as Diffusion Models are Trained (ICML 2025 · similarity 0.92)
- Swing-by Dynamics in Concept Learning and Compositional Generalization (ICLR 2025 · similarity 0.91)
- Improving Compositional Generation with Diffusion Models Using Lift Scores (ICML 2025 · similarity 0.90)
- How Diffusion Models Learn to Factorize and Compose (NeurIPS 2024 · similarity 0.90)
- Position: Compositional Generative Modeling: A Single Model is Not All You Need (ICML 2024 · similarity 0.90)
- CoInD: Enabling Logical Compositions in Diffusion Models (ICLR 2025 · similarity 0.89)
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.
