Diffusion Alignment Geometry
Methods for keeping diffusion models near natural image structure during preference optimization.
5 papers
Where this stands
The written synthesis of this thread is for subscribers. Subscribe.
How this thread developed
Oct 2026 · The University of Texas at Austin, Google
MEND selectively moves flow-model samples when reward gains justify distance
Introduces MEND, a RL method for aligning flow models with rewards without KL penalties, by selectively moving samples when reward gains justify distance.
2 further papers
Oct 2026 · Max Planck Institute for Informatics, Sapienza University of Rome
Adaptive steering fields control image generation while preserving structure
Replaces fixed steering vectors with state-dependent vector fields that adapt at every denoising step for structure-preserving image control.
Oct 2026 · The Hong Kong University of Science and Technology
Repelling crossing trajectories makes flow matching smoother and easier to sample
Repelling crossing trajectories in flow matching keeps trajectories near natural data structure, akin to geometric alignment
3 of 5 papers shown