MixFlow training method improves diffusion models by addressing exposure bias

Researchers have introduced MixFlow, a novel training approach designed to mitigate exposure bias in diffusion models. This method leverages "slowed interpolation mixtures" derived from the Slow Flow phenomenon, where ground-truth interpolations closer to generated data correspond to higher-noise timesteps. Experiments on class-conditional image generation, including SiT, REPA, and RAE models, demonstrate MixFlow's effectiveness. Notably, MixFlow applied to RAE models achieved strong generation results on ImageNet, with FID scores as low as 1.10 with guidance at both 256x256 and 512x512 resolutions. AI

IMPACT This new training method could lead to higher-quality image generation from diffusion models, potentially improving applications in creative fields and AI-assisted design.

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

Read on arXiv cs.AI →

• arXiv
• diffusion models
• Hui Li
• ImageNet
• MixFlow
• RAE
• Slow Flow

• paper
• model release

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