OmniColor: A Unified Framework for Multi-modal Lineart Colorization (ECCV 2026)

机器之心 · wechat · 2026-08-27

Accepted by ECCV 2026, the paper OmniColor proposes a unified multi-modal lineart colorization framework to handle the alignment and conflict of mixed control signals (e.g., text, color hints, reference images, history frames) in animation production. The core innovation categorizes control signals into two types: spatial alignment conditions (lineart, color hints, recent frames) and semantic reference conditions (text, identity reference, long-term frames). The model uses a dual-encoder strategy for spatial conditions, introduces a TRE module to eliminate redundancy in history frames, and an AS-Gate module to dynamically handle conflicts. Experiments show that the method outperforms existing approaches in quality, control precision, and sequence consistency, supporting flexible control combinations in real production workflows.

Original post →

More from Multimodal

Multimodal channel →