Model-Aware Schedules Improve Generation via Fiberwise Optimal Transport
Tailoring image generation schedules to match how each model actually learns
Researchers developed a way to customize the step-by-step schedules that guide image-generation AI models like DALL-E, by measuring how well each model predicts at different noise levels. The method produced a 38.6% improvement in image quality for one leading approach on standard benchmarks, and surprisingly, the optimal schedules followed similar patterns across different models and training setups.
Current image generators use one-size-fits-all schedules that don't account for how individual models actually perform. This work lets you extract better results from existing models without retraining them from scratch—the improvement template even works when frozen and applied to new models, potentially making high-quality image generation faster and cheaper across the board.