Vision-OPD: A Self-Distillation Method to Enable Multimodal Large Models to 'See Details Clearly'

Multimodal Large Language Models (MLLMs) have made significant progress in image understanding tasks, but fine-grained visual understanding still faces challenges: it is difficult to locate small yet critical visual evidence. The Vision-OPD framework, published in May 2026, uses a region-to-global self-distillation mechanism to enhance the model's ability to focus on fine-grained evidence under full-image input without relying on external teacher models or labeled data. Its core is to transfer the 'cropping advantage' of the model on cropped images to full-image reasoning.

The research team observed the 'region-to-global perception gap': when the same MLLM is input with a cropped image centered on the evidence, the accuracy of fine-grained question answering is much higher than when input with the complete image. This indicates that the model does not lack the ability to recognize local details, but rather struggles to focus on relevant evidence regions in the full image—i.e., 'can see details, but can't find where to look'. This insight points to the solution: transferring the cropping advantage to full-image reasoning.

Core Idea

Vision-OPD (Vision On-Policy Distillation) centers on distilling the model's own superior regional perception ability on cropped images into the full-image strategy, featuring self-distillation, online policy, no need for labels, and no additional tools during inference.

Teacher-Student Architecture

• Teacher Strategy: Input with evidence-centered cropped images, focusing on fine-grained features for more accurate token-level predictions.
• Student Strategy: Input with complete images (actual deployment scenario), aiming to learn the teacher's prediction distribution.

Distillation Process

1. The student generates a reasoning trajectory on the full image;
2. Calculate the difference in next-token probability distribution between the teacher (cropped image) and the student (full image);
3. Minimize this difference to let the student imitate the teacher's attention pattern;
4. End-to-end differentiable and trained via backpropagation.

Vision-OPD performs excellently on multiple fine-grained visual understanding benchmarks:

• Comparable to or even better than larger-scale open-source/closed-source models;
• Without inference tools (e.g., visual zoom), it can compete with tool-required agentic methods (e.g., Thinking-with-Images);
• Consistent improvement across MLLMs of different scales, with good generalization.

Advantages

1. No external resources needed: Does not rely on external teachers, labeled data, or reward models, reducing deployment costs;
2. Zero inference overhead: After training, only full-image input is required with no additional operations;
3. Strong versatility: Applicable to various MLLMs, not dependent on specific architectures or tasks.

Limitations

1. Depends on cropping quality: Teacher performance is affected by the cropping strategy;
2. Training complexity: Needs to maintain two strategies (teacher and student), and coordinating their interaction increases implementation difficulty.

Traditional methods usually rely on:

• High-resolution input: High computational cost;
• External teacher models: Increased dependencies and costs;
• Tools during inference: Increased latency;
• Labeled data: High cost.

Vision-OPD avoids all external dependencies and achieves similar or even better results through self-distillation.

Vision-OPD provides a concise and efficient solution for fine-grained visual understanding of MLLMs. Its core insight (the model already has fine-grained capabilities; the key is positioning) offers new ideas for domain research. In fields requiring high attention to details such as medical imaging, industrial inspection, and autonomous driving, such methods with zero inference overhead have important practical value.

References:

• Paper URL: http://arxiv.org/abs/2605.18740v1
• Publication date: May 18, 2026