Research map / Image generation / Fast diffusion sampling
RayFlow: Instance-Aware Diffusion Acceleration via Adaptive Flow Trajectories
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 RayFlow, a diffusion acceleration framework that routes each sample along an instance-specific flow trajectory toward a pretrained target mean, plus a Time Sampler that uses Stochastic Stein Discrepancy-based importance sampling over timesteps, enabling 1–8 step text-to-image generation.
Problem
Diffusion models generate slowly; existing few-step distillation and rectified-flow style acceleration methods lose sample quality, controllability, or require complex training.
Approach
Each sample follows a unique diffusion path toward an instance-specific target Gaussian N(ε_µ, σ²I) whose mean is a pretrained noise expectation, rather than a shared standard Gaussian, and a learned Time Sampler (SSD particles plus a neural network) gives importance-sampled timesteps for training.
Key results
- SD15-Ray reaches FID 5.10 at 1 step and 4.69 at 8 steps on COCO-5k, versus 19.8 for the SD15 base at 1 step (Table 1).
- SDXL-Ray achieves FID 4.15 at 1 step and 3.67 at 8 steps on COCO-5k, edging out SDXL-DMD2 (4.19 and 3.71) at those settings, while SDXL-PeRFlow has the best single-step aesthetic score among baselines (Table 1).
- Removing Time Sampler from SDXL on COCO-5k raises FID from 4.0 to 4.8 (Time Sampler table).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use RayFlow instead of ReFlow for SD15 | SD15 · COCO-5k, 2 steps | FID: 5.92 -> 5.10 | Helped |
| Use RayFlow instead of ReFlow for SDXL | SDXL · COCO-5k, 2 steps | FID: 4.95 -> 4.15 | Helped |
| Use RayFlow instead of PixArt-δ | PixArt · COCO-5k, 2 steps | FID: 4.76 -> 4.78 | Hurt |
| Add the RayFlow target mean (ablation) | SD15-Ray · COCO-5k, 2 steps | FID: 5.89 -> 5.10 | Helped |
| Add the Time Sampler (ablation) | SD15-Ray · COCO-5k, 2 steps | FID: 5.79 -> 5.10 | Helped |
| Add the RayFlow target mean (ablation) | SDXL-Ray · COCO-5k, 2 steps | FID: 4.85 -> 4.05 | Helped |
| Add the Time Sampler (ablation) | SDXL-Ray · COCO-5k, 2 steps | FID: 4.77 -> 4.05 | Helped |
| Add the RayFlow target mean (ablation) | PixArt-Ray · COCO-5k, 2 steps | FID: 5.50 -> 4.78 | Helped |
| Add the Time Sampler (ablation) | PixArt-Ray · COCO-5k, 2 steps | FID: 5.46 -> 4.78 | Helped |
| Add the Time Sampler (ablation) | SDXL-Ray · COCO-5k | FID: 4.8 -> 4.0 | Helped |
Surprising findings
- Single-step RayFlow outputs match or beat multi-step baselines in aesthetics on several models, though the reported gains over strong baselines are small (often within reported std).
Limitations
- Gains over strong baselines are small and often within reported standard deviations
- Target-mean determination is left for future work; the choice of sigma is sensitive (best at 0.3)
- Theoretical claim of path-probability maximization is tied to specific assumptions and only evaluated on images
Experimental setup
Datasets: LAION COYO COCO-5k ImageNet CIFAR-10 CIFAR-100
Benchmarks: COCO-5k ImageNet CIFAR-10 CIFAR-100
Metrics: FID CLIP score aesthetic score ImageReward PickScore
Baselines: SD15-LCM SD15-TCD Hyper-SD SDXL-Turbo SDXL-Lightning SDXL-DMD2 PeRFlow PixArt-δ
Builds on: rectified flow rectified diffusion latent consistency models DDPM flow matching
Closest papers by meaning
- A Simple Early Exiting Framework for Accelerated Sampling in Diffusion Models (ICML 2024 · similarity 0.89)
- Morse: Dual-Sampling for Lossless Acceleration of Diffusion Models (ICML 2025 · similarity 0.88)
- Test-time Scaling of Diffusions with Flow Maps (NeurIPS 2026 · similarity 0.88)
- Faster Diffusion: Rethinking the Role of the Encoder for Diffusion Model Inference (NeurIPS 2024 · similarity 0.87)
- Accelerating Rectified Flow Models via Trajectory-Aware Caching (NeurIPS 2026 · similarity 0.87)
- Shortcutting Pre-trained Flow Matching Diffusion Models is Almost Free Lunch (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.