InstructBLIP: Towards General-purpose Vision-Language Models with Instruction Tuning (2305.06500v2)

Abstract: Large-scale pre-training and instruction tuning have been successful at creating general-purpose LLMs with broad competence. However, building general-purpose vision-LLMs is challenging due to the rich input distributions and task diversity resulting from the additional visual input. Although vision-language pretraining has been widely studied, vision-language instruction tuning remains under-explored. In this paper, we conduct a systematic and comprehensive study on vision-language instruction tuning based on the pretrained BLIP-2 models. We gather 26 publicly available datasets, covering a wide variety of tasks and capabilities, and transform them into instruction tuning format. Additionally, we introduce an instruction-aware Query Transformer, which extracts informative features tailored to the given instruction. Trained on 13 held-in datasets, InstructBLIP attains state-of-the-art zero-shot performance across all 13 held-out datasets, substantially outperforming BLIP-2 and larger Flamingo models. Our models also lead to state-of-the-art performance when finetuned on individual downstream tasks (e.g., 90.7% accuracy on ScienceQA questions with image contexts). Furthermore, we qualitatively demonstrate the advantages of InstructBLIP over concurrent multimodal models. All InstructBLIP models are open-sourced at https://github.com/salesforce/LAVIS/tree/main/projects/instructblip.

• The paper demonstrates that instruction tuning on 26 datasets significantly improves zero-shot performance in vision-language tasks.
• It employs an instruction-aware Q-Former within a frozen BLIP-2 architecture to extract tailored visual features for diverse tasks.
• Ablation studies show that both instruction-aware feature extraction and data balancing are key to enhancing visual reasoning and multi-turn conversations.

InstructBLIP: Vision-LLMs Enhanced by Instruction Tuning

Introduction

The paper "InstructBLIP: Towards General-purpose Vision-LLMs with Instruction Tuning" addresses the complexity and diversity inherent in vision-language tasks through a novel framework leveraging instruction tuning. Vision-LLMs must handle inputs that are both textual and visual, making them significantly more challenging than traditional NLP tasks. This paper presents InstructBLIP, a model that utilizes instruction tuning to improve zero-shot generalization in vision-language tasks and demonstrates improved performance over previous models.

Vision-Language Instruction Tuning Methodology

The central innovation of InstructBLIP is its approach towards instruction tuning, which involves training a model on data formatted in natural language instructions. The paper employed a comprehensive collection of 26 publicly available datasets, encapsulating a wide array of tasks and then transformed them into an instruction tuning format.

Figure 1: Tasks and their corresponding datasets used for vision-language instruction tuning. The held-in datasets are indicated by yellow and the held-out datasets by white.

The architecture of InstructBLIP utilizes a pretrained BLIP-2 model, consisting of an image encoder, an LLM, and a Query Transformer (Q-Former). A key innovation is the instruction-aware Q-Former, which extracts visual features tailored to the specific instructions provided, optimizing the model's ability to follow varied instructions across differing tasks.

Figure 2: Model architecture of InstructBLIP. The Q-Former extracts instruction-aware visual features from the output embeddings of the frozen image encoder, and feeds the visual features as soft prompt input to the frozen LLM.

Experimental Results and Analysis

The InstructBLIP models achieved state-of-the-art performance in zero-shot evaluation tasks. Notably, InstructBLIP outperformed earlier models like Flamingo and BLIP-2 across a diverse range of datasets, such as video QA and visual reasoning tasks. These results underscore the efficacy of using instruction tuning to enhance the versatility and accuracy of vision-LLMs.

Figure 3: A few qualitative examples generated by our InstructBLIP Vicuna model. Here, a range of its diverse capabilities are demonstrated, including complex visual scene understanding and reasoning, knowledge-grounded image description, multi-turn visual conversation, etc.

An ablation paper was conducted to assess the contributions of instruction-aware visual features and data balancing strategies. The findings confirmed substantial improvements in performance due to these techniques, especially in datasets requiring intricate visual reasoning.

Instruction Tuning Versus Multitask Learning

The comparison between instruction tuning and multitask learning highlights a crucial insight: while multitask learning can perform well on seen tasks, instruction tuning proves to be superior in terms of generalizing to unseen tasks. This disparity indicates that the instructional format plays a critical role in enhancing a model's adaptability and understanding.

Figure 4: Comparison of instruction tuning and multitask training based on BLIP-2 FlanT5\textsubscript{XL}.

Practical Implications and Future Directions

InstructBLIP's architecture and methodology implies potential for wide application in various domains that require robust vision-language integration, from automated captioning to complex visual reasoning in real-time settings. The instruction-aware approach could expand further into dynamic, context-aware AI systems for even broader task generalization.

As research progresses, an exciting area of future development lies in further refining instruction tuning within multimodal settings, possibly integrating more sophisticated feedback mechanisms and broader datasets that encapsulate highly diverse contexts.

Conclusion

InstructBLIP represents a notable advancement in the field of vision-LLMs, demonstrating the power of instruction tuning to navigate the complex landscape of these tasks. By open-sourcing their models, the authors have provided a valuable resource for future research, encouraging the development of general-purpose multimodal AI systems that possess enhanced understanding and versatility.