EigDiff: Reducing Diffusion Training Data via CLIP Latent PCA
October 16, 2026
EigDiff reduces diffusion model training dataset requirements by selecting representative samples using principal component analysis in CLIP latent space. By using eigenvectors as reference points to pick both exemplars and diverse images, the method achieves performance parity with significantly smaller datasets.
HOW THIS AFFECTS YOU
●
builderYou may be able to significantly lower compute and data costs when fine-tuning diffusion models.
●
researcherThis offers a method to study data redundancy in diffusion training through latent space geometry.