Open access
2026
Practical Consequences of FID–CLIP Disagreement in Low-Data Diffusion Model Fine-Tuning
This paper examines when the Fréchet Inception Distance (FID) and a text–image similarity score based on Contrastive Language–Image Pre-training (CLIP) disagree in the ranking of such strategies, and gives initial evidence that these patterns are not specific to one architecture.
Weijun Zhou, Cuilian Zhang, Li-Xiang Song
· IEEE Access · 0 citations