New Adaptive Manifold Guidance improves generative models

Researchers have introduced Adaptive Manifold Guidance (AdaMaG), a new method for improving diffusion and flow-based generative models used in visual synthesis. AdaMaG addresses limitations in existing guidance techniques like Classifier-Free Guidance (CFG) by analyzing guidance through the continuity equation. This analysis reveals that guidance can be decomposed into a divergence term and a score-parallel term, and AdaMaG targets both to prevent samples from deviating from the learned manifold, especially under strong guidance. The method aims to enhance realism, reduce artifacts, and control desaturation in generated images without increasing inference costs. AI

IMPACT This new guidance method could lead to higher quality and more controllable image generation from diffusion and flow-based models.

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

Read on arXiv cs.CV →

• Adaptive Manifold Guidance
• arXiv
• Classifier-Free Guidance
• DagsHub
• Diffusion models
• Flow-based generative models
• Hugging Face
• Parsa Esmati
• Probability-Conserving Flow Guidance

• paper
• model release

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