Generative Models 5
- Your Loss Converged. The Model Learned Two Numbers.
- How to Actually Report FID: A Single-GPU Reproducibility Checklist
- Can DiffAugment Break the FID-160 Ceiling? A Limited-Data Remedy on a 2,490-Image CLIP-GAN
- Testing a Hypothesis Cheaply: Weakening the Discriminator Did Not Help in Four Runs
- MS-CLIP-GAN Architecture: How a CLIP-Guided Multi-Stage GAN Is Wired