LACON: Training Text-to-Image Model from Uncurated Data

TL;DR

LACON is a novel training framework for text-to-image models that leverages uncurated data by explicitly modeling quality signals, resulting in improved generation quality without data filtering.

Contribution

LACON introduces a method to utilize uncurated data by re-purposing quality signals as explicit labels, enhancing model performance over traditional filtered-data approaches.

Findings

LACON outperforms baselines trained on filtered data in image generation quality.

Explicit modeling of data quality boundaries improves the utilization of raw data.

Using uncurated data with LACON yields significant performance gains without additional compute.

Abstract

The success of modern text-to-image generation is largely attributed to massive, high-quality datasets. Currently, these datasets are curated through a filter-first paradigm that aggressively discards low-quality raw data based on the assumption that it is detrimental to model performance. Is the discarded bad data truly useless, or does it hold untapped potential? In this work, we critically re-examine this question. We propose LACON (Labeling-and-Conditioning), a novel training framework that exploits the underlying uncurated data distribution. Instead of filtering, LACON re-purposes quality signals, such as aesthetic scores and watermark probabilities as explicit, quantitative condition labels. The generative model is then trained to learn the full spectrum of data quality, from bad to good. By learning the explicit boundary between high- and low-quality content, LACON achieves superior generation quality compared to baselines trained only on filtered data using the same compute budget, proving the significant value of uncurated data.