Research map / Image generation / Fast diffusion sampling
Particle Guidance: non-I.I.D. Diverse Sampling with Diffusion Models
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes particle guidance, which adds a joint pairwise-similarity potential to diffusion reverse SDE/ODE so a set of samples is generated jointly to be diverse, with fixed and learned potential variants.
Problem
Generative models are usually sampled independently many times to get a diverse set, which is sample-inefficient; diversity of finite sets from diffusion models is not controlled.
Approach
Extend diffusion sampling to jointly sample a set of particles with a time-evolving permutation-invariant potential (e.g., negative pairwise RBF similarity) added as a gradient term to each particle's drift, plus a learned potential trained via a Doob h-transform objective to reach a target joint distribution.
Key results
- On GEOM-DRUGS, torsional diffusion with particle guidance raises median recall coverage from 80.0 to 82.6 and reduces median precision AMR from 0.729 to 0.594 (about 19%).
- Particle guidance achieves lower in-batch similarity than I.I.D. sampling at matched CLIP/Aesthetic scores for Stable Diffusion v1.5 on COCO prompts.
- Feature-space (DINO) particle guidance yields lower in-batch similarity than pixel/latent-space kernel, and SVGD produces blurry images.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Apply particle guidance to Torsional Diffusion | Torsional Diffusion · GEOM-DRUGS test | Recall AMR (median): 0.565 -> 0.520 | Helped |
| Apply particle guidance to Torsional Diffusion | Torsional Diffusion · GEOM-DRUGS test | Precision AMR (median): 0.729 -> 0.594 | Helped |
| Apply particle guidance to Torsional Diffusion | Torsional Diffusion · GEOM-DRUGS test | Recall coverage (median): 80.0 -> 82.6 | Helped |
| Apply particle guidance to Torsional Diffusion | Torsional Diffusion · GEOM-DRUGS test | Precision coverage (median): 56.9 -> 78.1 | Helped |
| Use particle guidance with a similarity kernel in DINO feature space instead of I.I.D. sampling | Stable Diffusion v1.5 · COCO 2014 validation | In-batch similarity | Helped |
| Apply particle guidance in DINO feature space rather than downsampled pixel space | Stable Diffusion v1.5 · COCO 2014 validation | In-batch similarity | Helped |
Surprising findings
- SVGD fails to promote diversity in image generation and yields blurry samples.
Limitations
- Arbitrary time-evolving potentials do not preserve marginal distributions, so the joint distribution is hard to characterize.
- Fixed potentials make optimality of the resulting joint distribution hard to argue about.
- Kernel choice (domain-specific) must be set by the user.
Experimental setup
Datasets: LAION-5B COCO 2014 GEOM-DRUGS
Benchmarks: COCO 2014 validation GEOM-DRUGS test
Metrics: in-batch similarity CLIP score Aesthetic score FID coverage average minimum RMSD
Baselines: I.I.D. sampling SVGD RDKit ETKDG OMEGA GeoMol GeoDiff Torsional Diffusion
Builds on: Stable Diffusion Torsional Diffusion classifier guidance Stein variational gradient descent metadynamics
Closest papers by meaning
- Greed is Good: A Unifying Perspective on Guided Generation (NeurIPS 2025 · similarity 0.89)
- Theoretical insights for diffusion guidance: A case study for Gaussian mixture models (ICML 2024 · similarity 0.88)
- Derivative-Free Guidance in Continuous and Discrete Diffusion Models with Soft Value-based Decoding (NeurIPS 2025 · similarity 0.88)
- Tortoise and Hare Guidance: Accelerating Diffusion Model Inference with Multirate Integration (NeurIPS 2025 · similarity 0.88)
- Unlocking Guidance for Discrete State-Space Diffusion and Flow Models (ICLR 2025 · similarity 0.88)
- Flow Density Control: Generative Optimization Beyond Entropy-Regularized Fine-Tuning (NeurIPS 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.