Research map / Image generation / Fast diffusion sampling
Logical Guidance for the Exact Composition of Diffusion Models
Francesco Alesiani, Jonathan H Warrell, Tanja Bien, Henrik Christiansen, Matheus Ferraz, Mathias Niepert
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 |
|---|---|---|---|
| use probability-weighted negation instead of constant-weight negation | LogDiff · CelebA (NOT) | FID: 19.7 -> 12.1 | Improved |
| use probability-dependent weights instead of constant weights for negation | LogDiff · CMNIST (NOT, higher guidance scale) | entropy / diversity | Improved |
| add adaptive repulsive guidance (A and not B) (ablation) | LogDiff · CMNIST (AND) | conformity score (%): 80.4 -> 83.6 | Improved |
| add adaptive repulsive guidance (A and not B) (ablation) | LogDiff · Shapes3D (AND) | conformity score (%): 82.4 -> 86.4 | Improved |
| add adaptive repulsive guidance (A and not B) (ablation) | LogDiff · CMNIST (OR-ME) | conformity score (%): 98.0 -> 97.9 | No clear effect |
| add adaptive repulsive guidance (A and not B) (ablation) | LogDiff · Shapes3D (N=2..5 operators) | conformity score (%): 88.8 -> 91.8 (N=2); 87.6 -> 89.9 (N=5) | Improved |
| add adaptive repulsive guidance with moderate weight (ablation) | EDM2-XXL on ImageNet-512 · ImageNet-512 | FID | Improved |
| use LogDiff composition instead of Skreta et al. (2025b) | diffusion composition · CMNIST and Shapes3D AND | conformity score (%): 26.2 -> 80.4 (CMNIST); 30.1 -> 82.4 (Shapes3D) | Improved |
| use LogDiff negation instead of Liu et al. (2022) constant-weight negation | LogDiff · CMNIST and Shapes3D (NOT) | conformity score (%): 99.5 -> 96.5 (CMNIST); 98.7 -> 96.4 (Shapes3D) | Hurt |
| add Feynman-Kac Correction to LogDiff (ablation) | LogDiff · GRM5-RRM1 A AND B | product docking score (A*B): 73.20 -> 80.89 | Improved |
| add Feynman-Kac Correction to LogDiff (ablation) | LogDiff · GRM5-RRM1 A OR B | product docking score (A*B): 73.91 -> 73.51 | Hurt |
| add Feynman-Kac Correction to LogDiff (ablation) | LogDiff · GRM5-RRM1 molecule diversity and validity | diversity: 0.89 -> 0.73 (AND) | Hurt |
Closest papers by meaning
- Derivative-Free Guidance in Continuous and Discrete Diffusion Models with Soft Value-based Decoding (NeurIPS 2025 · similarity 0.88)
- Simple Guidance Mechanisms for Discrete Diffusion Models (ICLR 2025 · similarity 0.88)
- Token Perturbation Guidance for Diffusion Models (NeurIPS 2025 · similarity 0.87)
- The Superposition of Diffusion Models Using the Itô Density Estimator (ICLR 2025 · similarity 0.87)
- Unlocking Guidance for Discrete State-Space Diffusion and Flow Models (ICLR 2025 · similarity 0.87)
- Training-Free Guidance Beyond Differentiability: Scalable Path Steering with Tree Search in Diffusion and Flow Models (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.
