Vision as LoRA
This addresses the structural complexity and computational overhead in MLLMs for AI researchers and practitioners, though it is incremental as it builds on existing LoRA and MLLM methods.
The paper tackles the problem of transforming large language models (LLMs) into multimodal large language models (MLLMs) by introducing Vision as LoRA (VoRA), which integrates vision-specific LoRA layers directly into the LLM to eliminate external vision modules, resulting in comparable performance to conventional encode-based MLLMs with additional pre-training data.
We introduce Vision as LoRA (VoRA), a novel paradigm for transforming an LLM into an MLLM. Unlike prevalent MLLM architectures that rely on external vision modules for vision encoding, VoRA internalizes visual capabilities by integrating vision-specific LoRA layers directly into the LLM. This design allows the added parameters to be seamlessly merged into the LLM during inference, eliminating structural complexity and minimizing computational overhead. Moreover, inheriting the LLM's ability of handling flexible context, VoRA can process inputs at arbitrary resolutions. To further strengthen VoRA's visual capabilities, we introduce a block-wise distillation method that transfers visual priors from a pre-trained ViT into the LoRA layers, effectively accelerating training by injecting visual knowledge. Additionally, we apply bi-directional attention masks to better capture the context information of an image. We successfully demonstrate that with additional pre-training data, VoRA can perform comparably with conventional encode-based MLLMs. All training data, codes, and model weights will be released at https://github.com/Hon-Wong/VoRA.