Zooming without Zooming: Region-to-Image Distillation for Fine-Grained Multimodal Perception (2026)

Free

Converts inference-time zooming into training-time primitive — teaches MLLMs fine-grained perception in single forward pass; introduces ZoomBench (845 VQA across 6 perceptual dimensions); SOTA on fine-grained benchmarks

FreeFree tier

Inputs: image, textOutputs: text

Type

Open Source

Visit Website GitHub (40 stars)

About Zooming without Zooming: Region-to-Image Distillation for Fine-Grained Multimodal Perception (2026)

Region-to-Image Distillation transforms iterative zooming from an inference-time tool into a training-time primitive, allowing multimodal large language models (MLLMs) to achieve fine-grained perception in a single forward pass. The method uses strong teacher models to generate high-quality VQA data on micro-cropped regions, then distills this region-grounded supervision back to the full image. The authors also introduce ZoomBench, a hybrid-annotated benchmark of 845 VQA data spanning six fine-grained perceptual dimensions, along with a dual-view protocol to quantify the global-regional 'zooming gap'. The approach achieves state-of-the-art performance on multiple fine-grained perception benchmarks and also improves general multimodal cognition on visual reasoning and GUI agent tasks. Code is available on GitHub.

Key Features

Transforms zooming from inference-time tool into training-time primitive

Enables fine-grained perception in a single forward pass of an MLLM

Uses teacher models to generate high-quality VQA data on micro-cropped regions

Distills region-grounded supervision back to the full image

Introduces ZoomBench: 845 VQA data spanning 6 fine-grained perceptual dimensions

Dual-view protocol quantifies global-regional zooming gap

Improves general multimodal cognition on visual reasoning and GUI agents

Pros & Cons

Pros

Reduces latency compared to iterative zooming methods by eliminating repeated tool calls and re-encoding
Achieves state-of-the-art performance on fine-grained perception benchmarks
Improves general multimodal cognition beyond fine-grained tasks
Open-source code and benchmark provided

Cons

May not fully replace iterative zooming for extremely complex scenes where iterative reasoning is beneficial
Training requires access to strong teacher models and high-quality region-annotated data
Performance depends on the quality of the teacher model used for distillation

Best For

Fine-grained visual perception in multimodal large language modelsVisual question answering (VQA) with region-level detailVisual reasoning tasks requiring global and local contextGUI agent interaction and understanding

FAQ

What is Region-to-Image Distillation?

It is a method that transforms the iterative zooming process used during inference into a training-time primitive, allowing MLLMs to internalize the benefits of agentic zooming into a single forward pass.

What is ZoomBench?

ZoomBench is a hybrid-annotated benchmark of 845 VQA data points spanning six fine-grained perceptual dimensions, with a dual-view protocol to quantify the global-regional 'zooming gap'.

How does Region-to-Image Distillation work?

First, strong teacher models generate high-quality VQA data by zooming into micro-cropped regions. Then, this region-grounded supervision is distilled back to the full image, so the student model learns fine-grained perception without needing tool use during inference.

What are the main results?

The method achieves leading performance across multiple fine-grained perception benchmarks and also improves general multimodal cognition on benchmarks such as visual reasoning and GUI agents.