PoGDiff: Product-of-Gaussians Diffusion Models for Imbalanced Text-to-Image Generation

Ziyan Wang, Sizhe Wei, Xiaoming Huo, Hao Wang · 2025

Read PoGDiff: Product-of-Gaussians Diffusion Models for Imbalanced Text-to-Image Generation by Ziyan Wang, Sizhe Wei, Xiaoming Huo, Hao Wang (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2502-08106.txt.

Description

Ziyan Wang, Sizhe Wei, Xiaoming Huo, and Hao Wang proposed PoGDiff in twenty twenty-five, a fine-tuning method that keeps image generators fair under skewed training data. Diffusion models lose quality on rare categories when image-text pairs over-represent common ones. Standard fine-tuning chases the ground-truth distribution directly, which bakes the skew in.

PoGDiff swaps that target for a product of Gaussians. The product blends the true target with the model's own prediction for a nearby text prompt, pulling rare and common categories toward shared ground. Tests on real data sets show better accuracy and image quality for under-represented content. The method is general, meant to apply across diffusion models facing imbalanced data, from product photos to medical images.

Catalog record

CREATORZiyan Wang, Sizhe Wei, Xiaoming Huo, Hao Wang
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE92 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2502-08106

Browse the full library · About the Tower of Babel