Research map / Image generation / Fast diffusion sampling
Divide-and-Denoise: A Game-Theoretic Method for Fairly Composing Diffusion Models
Abhi Gupta, Polina Barabanshchikova, Vikas K Garg, Samuel Kaski, Tommi 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 |
|---|---|---|---|
| add proportional fairness constraints (ablation) | Divide-and-Denoise, 2 models · Multi-object SD 2.0 | GenEval %images: 87.00% -> 88.50% | Improved |
| add proportional fairness constraints (ablation) | Divide-and-Denoise, 3 models · Multi-object SD 2.0 | GenEval %images: 51.75% -> 59.50% | Improved |
| add proportional fairness constraints (ablation) | Divide-and-Denoise, 3 models · Multi-object SD 2.0 | envy-free violation rate: 0.356 -> 0.017 | Improved |
| add proportional fairness constraints (ablation) | Divide-and-Denoise, 2 models · Multi-object SD 2.0 | CLIP min: 21.59 -> 21.53 | No clear effect |
| learned dynamic allocation instead of fixed vertical stripes (ablation) | Divide-and-Denoise vs MultiDiffusion, 3 models · Multi-object SD 2.0 | GenEval %images: 14.00% -> 59.50% | Improved |
| learned allocation instead of fixed stripes | Divide-and-Denoise vs MultiDiffusion · Descriptive concepts | GenEval %images: 27.00% -> 55.75% | Improved |
| learned allocation instead of fixed stripes | Divide-and-Denoise vs MultiDiffusion · Conflicting concepts | VQA joint: 0.819 -> 0.905 | Improved |
| team of single-concept models instead of one multi-concept model | Divide-and-Denoise vs Multi-Concept Diffusion · Multi-object SD 2.0 (3 models) | GenEval %images: 14.75% -> 59.50% | Improved |
| use attention-based utilities instead of score-based (ablation) | Divide-and-Denoise · Stable Diffusion | utility noise | Improved |
| add Divide-and-Denoise on DiT | DiT class-conditioned models · ImageNet class pairs | CLIP joint: 28.56 -> 29.03 | Improved |
| very large beta_t (regularization on allocation) (ablation) | Divide-and-Denoise · design analysis | allocation | Hurt |
Closest papers by meaning
- CADS: Unleashing the Diversity of Diffusion Models through Condition-Annealed Sampling (ICLR 2024 · similarity 0.89)
- Denoising Diffusion Bridge Models (ICLR 2024 · similarity 0.89)
- Your Diffusion Model is Secretly a Noise Classifier and Benefits from Contrastive Training (NeurIPS 2024 · similarity 0.89)
- A Simple Early Exiting Framework for Accelerated Sampling in Diffusion Models (ICML 2024 · similarity 0.89)
- Don't Play Favorites: Minority Guidance for Diffusion Models (ICLR 2024 · similarity 0.89)
- Divide-and-Conquer Posterior Sampling for Denoising Diffusion priors (NeurIPS 2024 · 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.
