Get Started
Home
Topics
Search
Library
Research questionHow can text-to-image diffusion models use multiple candidates and continuous rewards beyond pairwise preferences?When several images are generated for one prompt, reducing their graded reward scores to a winner-loser comparison discards information about their relative quality. That loss can limit how effectively the diffusion model learns the intended preferences.
Alignment & Safety
Diffusion Models
Image Generation
Machine Learning
Latest papersRecent research connected to this question, newest first.Beyond Pairwise Preferences: Listwise Reward-Aware Alignment for Diffusion ModelsApplies to text-to-image diffusion models used for generation, compositional generation, and image editing. The source studies reward-aware listwise optimization using candidate-group reward scores and a fixed reference model, with experiments on SD1.5 and SDXL.research paper · Sep 4, 2026
Related questions
How can text-to-image diffusion models preserve prompt alignment without the extra sampling cost of classifier-free guidance?How can diffusion image and video generators be preference-aligned without inefficient training exploration or inference-time search?How can preference optimization align whole-image preferences with token-specific content across spatial and denoising coordinates?How can prompt embeddings be optimized during diffusion inference to balance aesthetics, prompt alignment, and resource use?