Multimodal Instruction-Guided Image Editing Methods

With the rapid advancement of Artificial Intelligence Generated Content (AIGC) technologies, image editing tasks have achieved significant breakthroughs in generation quality. However, existing methods still face two major challenges: (1) reliance on single-modal guidance struggles to meet complex editing demands, and (2) the scarcity of high-quality multimodal datasets hampers model training. This paper addresses these issues through two core contributions. First, we design a multimodal instruction dataset construction pipeline leveraging multimodal large language models (MLLMs) and customized image generation techniques. This framework includes dataset generation and an MLLM-based filtering mechanism, significantly reducing the cost of creating multimodal datasets. Second, we desigened an enhanced version method based on the existing multimodal editing model Omnigen, improving its performance on multimodal instruction-guided image editing task. This innovation resolves the model's tendency to overlook non-text instructions during multi-instruction joint editing, thereby advancing multimodal instruction-guided image manipulation capabilities.