Research map / Image generation / Fast diffusion sampling
DiverseFlow: Sample-Efficient Diverse Mode Coverage in Flows
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.
Introduces DiverseFlow, a training-free inference-time method that couples flow ODE trajectories through a determinantal point process (DPP) volume gradient to get more diverse samples under a fixed sampling budget.
Problem
Flow-based generative models need many independent samples to cover multiple target modes; obtaining diverse sets this way is sample-inefficient, and diverse source samples do not guarantee diverse target samples.
Approach
Couple the K sample trajectories of a flow ODE by adding a gradient of a DPP log-likelihood (determinant of a similarity kernel over predicted target samples), with a quality term, so that diversity is driven during sampling rather than via repeated sampling or source perturbation.
Key results
- On the 2D mixture, DiverseFlow finds 7.11 modes on average vs 5.64 for IID sampling with MB-OT at budget K=10.
- On ImageNet-256 with LFM at CFG 1, DiverseFlow raises recall from 0.52 to 0.52 at equal precision; at CFG 4 recall rises from 0.26 to 0.38 with precision 0.69 to 0.70.
- Particle Guidance reaches 8.8 modes out of 10 on a symmetric mixture while DiverseFlow discovers all 10 on average; Particle Guidance yields lower aesthetic scores on polysemous prompts.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add DiverseFlow to LFM with CFG 1.5 | LFM · ImageNet-256 | recall: 0.44 -> 0.47 | Helped |
| Add DiverseFlow to LFM with CFG 2 | LFM · ImageNet-256 | recall: 0.41 -> 0.46 | Helped |
| Add DiverseFlow to LFM with CFG 4 | LFM · ImageNet-256 | recall: 0.26 -> 0.38 | Helped |
| Add DiverseFlow to LFM with CFG 1 | LFM · ImageNet-256 | recall: 0.52 -> 0.52 | No clear effect |
| Add DiverseFlow to LFM with CFG 2 | LFM · ImageNet-256 | precision: 0.77 -> 0.76 | Hurt |
| Add DiverseFlow to MB-OT | MB-OT · Synthetic 2D Gaussian mixture, sampling budget 10 | average modes discovered: 5.64 -> 7.11 | Helped |
| Use DiverseFlow instead of IID sampling | Flow sampling · Symmetric uniform Gaussian mixture, sampling budget 10 | average modes discovered: 6.5 -> 10 | Helped |
| Use DiverseFlow instead of Particle Guidance | Flow sampling · Symmetric uniform Gaussian mixture, sampling budget 10 | average modes discovered: 8.8 -> 10 | Helped |
| Add DiverseFlow to flow matching | Flow matching formulations · Synthetic 2D Gaussian mixture | average modes discovered | Helped |
| Add DiverseFlow to LFM with classifier-free guidance | LFM · ImageNet-256 | precision: 0.62 -> 0.62; 0.69 -> 0.69; 0.77 -> 0.76; 0.69 -> 0.70 | Mixed |
Surprising findings
- Particle Guidance achieves high diversity scores but produces image artefacts that lower aesthetic quality on polysemous prompts.
- Diverse source samples do not necessarily yield diverse target samples under flow mapping.
Limitations
- Gradient-based coupling adds per-step cost via DPP gradient computation and depends on the quality of one-step target estimates.
- Diversity measured on feature-space distances requires a domain-specific feature extractor (e.g., ViT).
Experimental setup
Datasets: ImageNet CelebA-HQ
Benchmarks: ImageNet-256 precision/recall CelebA-HQ inpainting synthetic 2D mixture of Gaussians mode discovery polysemous prompt set
Metrics: precision recall number of modes discovered aesthetic score average pairwise similarity
Baselines: IID sampling Particle Guidance Rectified-Flow + MCG
Builds on: flow matching classifier-free guidance manifold constrained gradient particle guidance
Closest papers by meaning
- Diffusion Generative Flow Samplers: Improving learning signals through partial trajectory optimization (ICLR 2024 · similarity 0.88)
- Align Your Flow: Scaling Continuous-Time Flow Map Distillation (NeurIPS 2025 · similarity 0.88)
- Discovering Latent Graphs with GFlowNets for Diverse Conditional Image Generation (NeurIPS 2025 · similarity 0.87)
- Expanding Flow Maps (NeurIPS 2026 · similarity 0.87)
- Image Generation Diversity Issues and How to Tame Them (CVPR 2025 · similarity 0.87)
- Beyond Squared Error: Exploring Loss Design for Enhanced Training of Generative Flow Networks (ICLR 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.