Abstract
Recent diffusion-based video generation models have made significant progress in multi-reference image-conditioned video editing. However, existing methods still struggle to coordinate information from multiple visual sources accurately. We identify that instruction quality is a critical limiting factor: existing editing instructions lack explicit reference attribution, and general-purpose multimodal large language models (MLLMs) cannot generate them reliably. To address this, we propose ReBind, a systematic framework that introduces dense instructions with explicit reference relationships as the intermediate representation. Our key insight is embedding reference tokens at semantic positions to establish precise bindings between visual attributes and their sources. We develop ReBind-Instruct, a specialized MLLM trained through a two-stage scheme: Structured Instruction Learning teaches the model to generate reference-grounded instructions, while Reference Attribution Optimization uses reinforcement learning with tailored rewards to improve content completeness and attribution accuracy. We further develop ReBind-Edit, which leverages these dense instructions to perform precise multi-reference video editing through cross-attention binding. Extensive experiments show that ReBind substantially outperforms general-purpose MLLMs in instruction quality and achieves state-of-the-art performance among open-source methods, with ablations confirming that both instruction structure and quality directly impact editing performance.
Figure: Reward design for Reference Attribution Optimization. Our tailored rewards improve both content completeness and attribution accuracy in dense instruction generation.
