New RACER method improves diffusion model speed and quality

Researchers have developed a new method called RACER (Robust Adaptive Cost Efficient Routing) to improve the speed and quality of diffusion models. Unlike previous approaches that blindly trust forecasts for skipped denoising steps, RACER observes the agreement between forecasts to gauge reliability. This closed-loop controller dynamically adjusts its trust in forecasts, shrinking uncertain ones and refreshing features at critical steps. RACER has demonstrated improved performance across several diffusion models, including SD3.5-Large, FLUX.1-dev, Wan2.1-14B, and HunyuanVideo, on benchmarks like DrawBench, VBench, and COCO, while also enabling faster sampling at equal quality on SD3.5. AI

IMPACT Accelerates diffusion model sampling and improves quality, potentially leading to faster and more efficient image and video generation.

RANK_REASON The item is a research paper detailing a new method for diffusion models. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.LG →

• arXiv
• COCO
• diffusion models
• DrawBench
• FLUX.1-dev
• HunyuanVideo
• SD3.5
• SD3.5-Large
• Taylor
• VBench
• Wan2.1-14B

• paper
• infra

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →