Agentic data science with the Data Agent Kit
cloud.google.com·7h ago
AI Research
arXiv CS.AI·Sep 12·5 min read
Model-Aware Schedules Improve Generation via Fiberwise Optimal Transport
TL;DR
A new model-aware schedule construction using fiberwise optimal transport significantly enhances data generation in diffusion models. This approach achieved a 38.6% relative reduction in FID scores on CIFAR-10 compared to strong baselines.
✦ Why It Matters
Implement model-aware schedules in your diffusion models to enhance generation quality immediately.
Key Takeaways
- Adopt fiberwise optimal transport for better signal/noise management.
- Utilize closed-form solutions for optimal time allocation in models.
- Evaluate model performance using FID scores for benchmarking.
- Consider frozen analytic templates to retain improvements without retraining.
How It Works
The model-aware schedule construction uses fiberwise optimal transport to create compatible signal and noise decompositions, forming affine fibers. By calculating a fiberwise prediction risk, the method optimally allocates time along coefficient paths, enhancing data generation.