"T2I-ConBench"

"Continual post-training enables a single text-to-image (T2I) diffusion model to incrementally acquire new capabilities without maintaining separate models, but na\u00efve adaptation often causes catastrophic forgetting and weakens compositional generation. We observe the absence of a standardized evaluation protocol for continual post-training T2I diffusion model research. To address this gap, we introduce T2I-ConBench, a unified benchmark for continual post-training of T2I diffusion models. T2IConBench focuses on two practical scenarios, item customization and domain enhancement, and evaluates four key aspects: retention of pretrained generality, target-task adaptation, forgetting, and cross-task generalization. It combines automatic metrics, humanpreference modeling, and vision-language Q&A to provide a comprehensive assessment of continual post-training behavior. Building on this benchmark, we construct three realistic task streams with mixed granularity and evaluate ten representative methods. The results show that no single method performs best across all settings, even Joint learning is not uniformly optimal, targeted post-training can unlock pretrained capabilities, and cross-task generalization remains a major open challenge."