Research map / Image generation / Fast diffusion sampling
FlowTurbo: Towards Real-time Flow-Based Image Generation with Velocity Refiner
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 FlowTurbo, which accelerates sampling of flow-based image generators by training a lightweight velocity refiner (about 5% of parameters) plus a pseudo corrector and sample-aware compilation, reaching FID 2.12 on ImageNet at 100 ms/img.
Problem
Sampling from flow-based generative models is slow; existing samplers are classical numerical ODE solvers (Euler, Heun), and one-step distillation methods sacrifice multi-step flexibility and need costly training.
Approach
Flow-model velocity predictions are stable across sampling steps, so a lightweight refiner can regress the velocity offset from the previous step's velocity and replace the full predictor at selected steps; a pseudo corrector reuses the previous step's velocity to halve evaluations while keeping second-order convergence, and sample-aware compilation compiles the whole sampling loop with CFG.
Key results
- On ImageNet 256x256, FlowTurbo H8P9R5 reaches FID 2.12 at 100.3 ms/img versus 2.20 FID for 25-step Heun's SiT-XL at 250 ms/img.
- FlowTurbo H1P5R3 achieves FID 3.93 at 38 ms/img, which the authors describe as real-time generation.
- Acceleration of 53.1%-58.3% on class-conditional generation and 29.8%-38.5% on text-to-image generation relative to Heun's baselines at similar FID.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Replace selected velocity-predictor evaluations with a velocity refiner and use pseudo-corrector blocks | SiT-XL · ImageNet 256×256 | FID↓: 3.68 -> 3.63; 2.79 -> 2.69; 2.42 -> 2.22; 2.20 -> 2.12 | Helped |
| Replace Heun sampling with FlowTurbo blocks | SiT-XL · ImageNet 256×256 | latency: 89.4 -> 41.6; 117.8 -> 55.2; 154.8 -> 72.5; 240.6 -> 100.3 | Helped |
| Replace Heun sampling with FlowTurbo blocks | InstaFlow 2-RF · MS COCO 2017 512×512 | FID↓: 32.77 -> 32.48; 30.73 -> 30.19; 28.61 -> 28.60; 28.06 -> 27.60 | Mixed |
| Replace Heun sampling with FlowTurbo blocks | InstaFlow 2-RF · MS COCO 2017 512×512 | latency: 104.5 -> 68.4; 120.3 -> 84.5; 170.5 -> 104.8; 203.7 -> 137.0 | Helped |
| Add one velocity-refiner block (ablation) | FlowTurbo on SiT-XL · ImageNet 256×256 | FID↓: 4.42 -> 2.80 | Helped |
| Add one velocity-refiner block (ablation) | FlowTurbo on SiT-XL · ImageNet 256×256 | latency: +0.7% | No clear effect |
| Use sample-aware instead of model-level compilation (ablation) | FlowTurbo on SiT-XL · ImageNet 256×256 | latency: 41.0 -> 38.3 | Helped |
| Vary the training Δt range across reported intervals (ablation) | FlowTurbo on SiT-XL · ImageNet 256×256 | FID↓: 3.58; 3.55; 4.48; 3.36; 2.73; 2.65; 2.93; 2.62; 2.53; 2.54; 2.64; 2.89 | Mixed |
| Increase the number of velocity-refiner blocks from 1 to 3 (ablation) | FlowTurbo on SiT-XL · ImageNet 256×256 | FID↓: 2.80; 3.55; 7.62 | Mixed |
| Increase pseudo-corrector blocks while reducing Heun blocks (ablation) | FlowTurbo on SiT-XL · ImageNet 256×256 | latency: 62.0; 58.6; 55.2; 51.8; 48.4; 45.0; 41.6 | Helped |
| Increase pseudo-corrector blocks while reducing Heun blocks (ablation) | FlowTurbo on SiT-XL · ImageNet 256×256 | FID↓: 2.93; 2.60; 2.66; 2.78; 2.96; 3.21; 3.59 | Hurt |
| Vary the number of velocity refiners from 3 to 6 with H7P10 (ablation) | FlowTurbo on SiT-XL · ImageNet 256×256 | FID↓: 2.19; 2.12; 2.18; 2.15 | Mixed |
Surprising findings
- Velocity predictions of flow models are stable along the sampling trajectory, whereas diffusion epsilon predictions are not, which the authors did not find in diffusion models.
Limitations
- Relies on velocity stability, which the authors have not observed in diffusion-based models.
- Potential misuse for accelerating malicious content generation.
Experimental setup
Datasets: ImageNet-1K LAION (50K subset) MS COCO 2017
Benchmarks: ImageNet 256x256 FID 50K MS COCO 2017 FID 5K
Metrics: FID IS precision recall latency (ms/img) FLOPs
Baselines: Heun's method (SiT-XL, 25 steps) dopri5 adaptive solver DiT-XL ADM-G LDM-4-G StyleGAN-XL Mask-GIT SD2.1 SDXL PixArt-alpha PixArt-sigma
Builds on: Heun's method classifier-free guidance flow matching SiT InstaFlow
Closest papers by meaning
- Align Your Flow: Scaling Continuous-Time Flow Map Distillation (NeurIPS 2025 · similarity 0.88)
- Bellman Optimal Stepsize Straightening of Flow-Matching Models (ICLR 2024 · similarity 0.88)
- Simple ReFlow: Improved Techniques for Fast Flow Models (ICLR 2025 · similarity 0.88)
- Mean Flows for One-step Generative Modeling (NeurIPS 2025 · similarity 0.87)
- Self-Corrected Flow Distillation for Consistent One-Step and Few-Step Image Generation (AAAI 2025 · similarity 0.87)
- Consistent Flow Distillation for Text-to-3D Generation (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.