Research map / Image generation / Fast diffusion sampling
Gaussian Mixture Flow Matching 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 GMFlow, which predicts a per-pixel Gaussian mixture over flow velocity trained with a KL/NLL loss, plus analytic GM-SDE/ODE solvers and a probabilistic guidance scheme that reweights the GM to avoid CFG over-saturation.
Problem
Flow matching and diffusion models need many sampling steps because single-Gaussian denoising approximations and ODE discretization cause errors, and classifier-free guidance over-saturates colors.
Approach
Replace the single-Gaussian mean prediction with a Gaussian mixture over flow velocity trained by KL divergence; derive analytic GM denoising transitions and velocity fields for few-step SDE/ODE solvers, and apply CFG as reweighting of mixture components by a Gaussian mask rather than extrapolation.
Key results
- GM-ODE 2 reaches Precision 0.942 on ImageNet 256x256 in 6 sampling steps (K=8).
- GM-SDE 2 attains Precision 0.950 at 32 steps, reported as state-of-the-art in the paper.
- On CIFAR-10 with 10 NFE, GM-SDE 2 achieves FID 9.11 vs 13.80 for GMS (the best competitor listed).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use GM-SDE 2 instead of GM-SDE (ablation) | GMFlow · ImageNet, NFE=8 | FID: 6.96 (+3.53) -> 3.43 | Helped |
| Add spectral sampling (ablation) | GM-SDE · ImageNet, NFE=8 | FID: 8.98 (+2.02) -> 6.96 | Helped |
| Use probabilistic guidance instead of vanilla CFG (ablation) | GM-SDE · ImageNet, NFE=8 | Precision: 0.917 (−0.023) -> 0.940 | Helped |
| Use GM-SDE 2 instead of DPM++ 2M SDE (ablation) | GMFlow · ImageNet, NFE=8 | Precision: 0.912 (−0.027) -> 0.939 | Helped |
| Use transition loss rather than the original KL loss (λ=1.0) (ablation) | GM-SDE 2 · ImageNet, NFE=8 | FID: 4.49 (+1.06) -> 3.43 | Helped |
| Use GM-ODE sub-steps (ablation) | GM-ODE 2 · ImageNet, NFE=8 | FID: 7.47 (+4.70) -> 2.77 | Helped |
| Use the q̂(x₀|xₜ) conversion in second-order GM solvers (ablation) | GMFlow second-order GM solvers · 2D checkerboard | sample uniformity | Helped |
| Increase the number of mixture components (ablation) | GMFlow · 2D checkerboard, few-step sampling | sample quality | Helped |
| Increase mixture components beyond K=8 (ablation) | GM-SDE 2 · ImageNet, few-step sampling | Precision | Hurt |
| Add spectral sampling (ablation) | GMFlow · ImageNet training images (latents) | NLL (bits/dim): 0.224 -> 0.173 | Helped |
| Increase GM components from K=1 to K=16 (ablation) | GMFlow · ImageNet training images (latents) | NLL (bits/dim): 0.346 -> 0.224 | Helped |
| Use GM-SDE 2 instead of competitor methods | CIFAR-10 unconditional image generation · CIFAR-10 | FID: 9.11 vs 13.80; 4.16 vs 5.48; 3.79 vs 4.00; 3.76 vs 3.46 | Mixed |
Surprising findings
- Vanilla CFG with the GM mean shifted (ablation A3) causes severe precision drop and over-saturation, while probabilistic guidance avoids it.
- Precision of GM-SDE 2 declines beyond K=8 because spectral sampling introduces larger numerical errors with more components.
Limitations
- Pixel-wise factorization for high-dimensional images may not fully exploit GMFlow's potential.
- Spectral sampling is needed to recover spatial correlations, and its numerical errors grow with more GM components.
- Second-order GM solvers slightly underperform at higher guidance scales.
Experimental setup
Datasets: ImageNet CIFAR-10 2D checkerboard
Benchmarks: ImageNet 256x256 CIFAR-10
Metrics: FID Precision Recall Inception Score Saturation NLL
Baselines: DDPM DDIM/Euler DPM++ DEIS UniPC SA-Solver GMS SN-DDPM
Builds on: flow matching DiT DPM-Solver++ UniPC DDPM
Closest papers by meaning
- On the Closed-Form of Flow Matching: Generalization Does Not Arise from Target Stochasticity (NeurIPS 2025 · similarity 0.89)
- Improving Progressive Generation with Decomposable Flow Matching (NeurIPS 2025 · similarity 0.88)
- On the Guidance of Flow Matching (ICML 2025 · similarity 0.87)
- Coreset-Induced Conditional Velocity Flow Matching (NeurIPS 2026 · similarity 0.87)
- Discrete Flow Matching (NeurIPS 2024 · similarity 0.87)
- Flow matching achieves almost minimax optimal convergence (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.