OmniConsistency: Learning Style-Agnostic Consistency from Paired Stylization Data

Diffusion models have advanced image stylization significantly, yet two corechallenges persist: (1) maintaining consistent stylization in complex scenes,particularly identity, composition, and fine details, and (2) preventing styledegradation in image-to-image pipelines with style LoRAs. GPT-4o's exceptionalstylization consistency highlights the performance gap between open-sourcemethods and proprietary models. To bridge this gap, we propose\textbf{OmniConsistency}, a universal consistency plugin leveraging large-scaleDiffusion Transformers (DiTs). OmniConsistency contributes: (1) an in-contextconsistency learning framework trained on aligned image pairs for robustgeneralization; (2) a two-stage progressive learning strategy decoupling stylelearning from consistency preservation to mitigate style degradation; and (3) afully plug-and-play design compatible with arbitrary style LoRAs under the Fluxframework. Extensive experiments show that OmniConsistency significantlyenhances visual coherence and aesthetic quality, achieving performancecomparable to commercial state-of-the-art model GPT-4o.