Research map / Image generation / Fast diffusion sampling
Compositional Sculpting of Iterative Generative Processes
Timur Garipov, Sebastiaan De Peuter, Ge Yang, Vikas Garg, Samuel Kaski, Tommi S. Jaakkola
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 |
|---|---|---|---|
| compose two pre-trained GFlowNets with the harmonic mean operation | GFlowNet composition · 2D grid | probability mass overlap with both base modes | Improved |
| use contrast operation instead of harmonic mean | GFlowNet composition · 2D grid | modes of p1 modulated by p2 | Mixed |
| use parameterized contrast with alpha = 0.05 (ablation) | GFlowNet composition · 2D grid | suppression of p2 regions | Improved |
| condition on three labels instead of two (ablation) | GFlowNet composition · 2D grid | sharpness of composite | Improved |
| harmonic mean of SEH and SA GFlowNets | GFlowNet composition · molecule generation beta 32/96 | reward distribution | Improved |
| contrast operation for SEH and SA GFlowNets | GFlowNet composition · molecule generation beta 32/96 | reward distribution | Mixed |
| condition on SEH, QED and SA together | GFlowNet composition · molecule generation beta 32 | fraction of samples with all three properties high | Improved |
| condition on two labels (SEH and QED) only | GFlowNet composition · molecule generation beta 32 | reward in the unselected reward | No clear effect |
| repeat the same label (SEH, SEH, SEH) when conditioning | GFlowNet composition · molecule generation | distance to base distribution | Mixed |
| train classifier on terminal states first before non-terminal states (ablation) | Diffusion classifier · diffusion compositions | training noise | Improved |
| compose three colored-MNIST diffusion models by conditioning on y1, y2, y3 | Diffusion composition · colored MNIST | samples satisfying all constraints | Improved |
| omit conditioning on p2 in composition of y1=1, y2=1 | Diffusion composition · colored MNIST | excluded digits | Improved |
Closest papers by meaning
- Improving Compositional Generation with Diffusion Models Using Lift Scores (ICML 2025 · similarity 0.89)
- IterComp: Iterative Composition-Aware Feedback Learning from Model Gallery for Text-to-Image Generation (ICLR 2025 · similarity 0.88)
- How Diffusion Models Learn to Factorize and Compose (NeurIPS 2024 · similarity 0.87)
- Position: Compositional Generative Modeling: A Single Model is Not All You Need (ICML 2024 · similarity 0.87)
- How Compositional Generalization and Creativity Improve as Diffusion Models are Trained (ICML 2025 · similarity 0.87)
- Composition and Alignment of Diffusion Models using Constrained Learning (NeurIPS 2025 · similarity 0.87)
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.
