FlowTurbo: Accelerating Flow-Based Image Generation Models via Multi-Stage Refinement.

Area of Science:

• Artificial Intelligence
• Computer Vision
• Machine Learning

Background:

• Flow-based models offer competitive visual generation quality and inference speed.
• Efficient sampling methods for flow-based models are underexplored compared to diffusion models.
• Flow-matching enables straighter sampling trajectories beneficial for generation.

Purpose of the Study:

• To develop a framework, FlowTurbo, for accelerating the sampling process of flow-based generative models.
• To enhance both the sampling speed and visual quality of flow-based models.
• To introduce novel techniques for reducing inference time in generative tasks.

Main Methods:

• Proposed FlowTurbo framework utilizing a lightweight velocity refiner to estimate stable velocity outputs.
• Introduced a pseudo corrector and sample-aware compilation for further inference time reduction.
• Developed a multi-stage refinement technique splitting generation across resolutions for large models.
• Implemented a stage-aware deployment strategy to optimize latency and throughput.

Main Results:

• Achieved acceleration ratios of 53.1%–58.3% for class-conditional generation and 29.8%–38.5% for text-to-image generation.
• Reached FID of 2.12 (100 ms/img) and 3.93 (38 ms/img) on ImageNet, establishing new state-of-the-art real-time generation.
• Demonstrated ~50% speed improvement with SD 3.5 Large, achieving FID of 28.05 on NVIDIA 3090 GPU.

Conclusions:

• FlowTurbo effectively accelerates flow-based generative models without altering the multi-step sampling paradigm.
• The framework is applicable to various tasks like image editing and inpainting.
• FlowTurbo enables real-time, high-fidelity image generation, setting new benchmarks in the field.