Agentic data science with the Data Agent Kit
cloud.google.com·8h ago
AI Research
arXiv CS.AI·Aug 7·5 min read
PRISM: Distribution-Gated Flow Matching for Controllable Unpaired Image Translation
TL;DR
PRISM introduces a GAN-free flow-matching framework for unpaired image translation, utilizing learned per-feature gates to control content preservation and appearance changes. It outperforms existing methods on multiple benchmarks, achieving superior realism and structural fidelity.
✦ Why It Matters
Explore PRISM for advanced image translation tasks in your projects without needing extensive retraining.
Key Takeaways
- Implement PRISM for nuanced image transformations in your applications.
- Utilize learned per-feature gates for better content preservation.
- Evaluate performance using Inception FID and KID metrics.
- Apply PRISM in medical imaging for improved structural fidelity.
How It Works
PRISM replaces global control with learned per-feature gates that assess the distance of each feature from the target distribution. Features that are far from the target are allowed to change, while those that are consistent with the target are preserved.
This mechanism enables more precise control over the translation process, allowing for both structure-preserving and structure-changing translations.