Region-to-Image Distillation

• The paper demonstrates that training-time distillation from micro-cropped regions enables accurate fine-grained multimodal perception by compressing the zooming benefit into model parameters.
• The paper introduces a two-stage approach where a teacher generates micro-evidence from cropped regions, which is then grounded back to full-image inputs via an explicit overlay transformation.
• The paper reports significant empirical improvements by reducing the zooming gap and achieving faster single-pass inference compared to conventional tool-using pipelines.

Region-to-Image Distillation (R2I) is a training-time distillation procedure for fine-grained multimodal perception in which supervision is first generated on micro-cropped image regions and then transferred back to full-image inputs, with the goal of enabling “single-glance” perception without test-time tool use (Wei et al., 12 Feb 2026). In the formulation introduced by “Zooming without Zooming: Region-to-Image Distillation for Fine-Grained Multimodal Perception,” the method is designed for settings where decisive visual evidence is small, visually cluttered, and easily overwhelmed by global context; it is therefore positioned as an alternative to inference-time “Thinking-with-Images” pipelines that repeatedly crop, zoom, and re-encode regions of interest (Wei et al., 12 Feb 2026). The term should be distinguished from “Region-aware Knowledge Distillation ReKo,” which concerns image-to-image translation model compression rather than multimodal fine-grained perception (Zhang et al., 2022).

1. Conceptual basis and problem setting

R2I is motivated by a specific diagnosis of multimodal LLM (MLLM) failure: many fine-grained errors arise not from missing reasoning ability, but from inadequate access to micro-evidence in the original image (Wei et al., 12 Feb 2026). Recent “Thinking-with-Images” (TwI) systems such as DeepEyes, Thyme, and Mini-o3 improve accuracy by iteratively zooming and re-encoding image regions during inference, but this introduces latency and tool overhead because the pipeline requires repeated visual encoding passes and agentic loops. R2I asks whether the practical gain comes from the tool use itself or from improved access to small, decisive evidence.

The method’s answer is framed in terms of information neutrality. When a tool action does not add new external information, but merely exposes evidence already present in the image more effectively, the gain may be distillable into the model weights through training-time supervision. In that sense, R2I treats zooming as a training primitive rather than an inference primitive: the model is taught to associate a full-image view with an answer originally grounded in a micro-region. This suggests a broader view of distillation in which the object being distilled is not only a teacher model’s output distribution, but also the functional benefit of a visual tool action.

The paper also presents this idea more abstractly as “tool-action distillation.” Given an image $I$, a tool action $f(\cdot)$ produces $\widehat{I}$; a teacher then generates question–answer pairs $(Q,A) \sim \mathcal{T}(\widehat{I})$; finally, an inverse mapping $f^{-1}$ produces $(I,\widehat{Q},A)$ for student training (Wei et al., 12 Feb 2026). In the reported implementation, $f$ is zooming, and the inverse is realized by grounding the target region on the full image with a bounding-box overlay.

2. Region selection and micro-crop synthesis

The R2I pipeline has two stages: zoom-in synthesis on micro-crops and zoom-out distillation back to the full image (Wei et al., 12 Feb 2026). The construction begins from a raw image pool $\mathcal{D}_{\text{raw}}$ and a proposal function $\mathcal{P}(I)$ that generates candidate boxes $\{B_1, B_2, \dots, B_n\}$ using an object recognition and segmentation system. A target-region filter then keeps only small, evidence-hidden regions according to

$f(\cdot)$0

with a sparsity threshold $f(\cdot)$1, exemplified as $f(\cdot)$2 (Wei et al., 12 Feb 2026).

The object-centric character of these proposals is methodologically important. The retained regions are intended to cover at least one visible object and thus to be semantically meaningful rather than arbitrary patches. In the implementation appendix, the image sources are specified as high-resolution images, mostly above $f(\cdot)$3, curated from SA-1B, LAION, MetaCLIP, Visual Genome, CC12M, and STPLS3D. For datasets without object annotations, Qwen3-VL-235B is used to generate object inventories and SAM3 to obtain bounding boxes. Regions with area ratio below $f(\cdot)$4 are retained, cropped, and resized by $f(\cdot)$5 before synthetic VQA generation (Wei et al., 12 Feb 2026).

On each micro-crop $f(\cdot)$6, strong teacher models generate region-grounded questions and answers. The questions are required to be answerable from the crop alone and to focus on details that are difficult to resolve from the full image, including tiny text, symbols, subtle attributes, and fine counts. In the main method description, a teacher acts as a question generator to propose perception-centric questions $f(\cdot)$7, and teacher answer generators produce pseudo-labels $f(\cdot)$8 by majority voting. A triplet $f(\cdot)$9 is retained only when the answer generators reach high consensus, thereby reducing hallucination and excluding ambiguous or invalid questions (Wei et al., 12 Feb 2026).

The appendix provides the concrete synthesis stack. Qwen3-VL-235B is used as the question generator, while Qwen3-VL-235B and GLM-4.5V serve as two independent answer generators. Four responses are sampled per answer generator, yielding eight total answers, and a QA pair is retained only when the majority answer reaches strict consensus greater than $\widehat{I}$0. Difficulty filtering is then applied with Qwen3-VL-8B, discarding data that the smaller model can answer correctly more than twice in four trials. The resulting training set contains 74K samples (Wei et al., 12 Feb 2026).

3. Grounding transformation and student optimization

The second stage converts crop-grounded supervision into full-image supervision. This is necessary because a question that is unambiguous on a crop can become referentially ambiguous when transferred to the whole image. The paper resolves this with an explicit grounding transformation $\widehat{I}$1, described as an inverse of $\widehat{I}$2, which overlays the bounding box on the full image and appends a spatial constraint to the question (Wei et al., 12 Feb 2026). The resulting training triplet is $\widehat{I}$3, where $\widehat{I}$4 is the full image with an overlaid box and $\widehat{I}$5 is the region-anchored prompt.

The student is then trained to maximize expected reward on these synthetic triplets:

$\widehat{I}$6

Here, $\widehat{I}$7 denotes the student policy and $\widehat{I}$8 is a task reward (Wei et al., 12 Feb 2026). The appendix states that optimization is performed with reinforcement learning via DAPO in the EasyR1 framework rather than supervised fine-tuning.

The reward design is tiered. First, exact or symbolic matching is applied. Second, for numeric tasks a continuous counting reward is used:

$\widehat{I}$9

for non-zero ground truth, and

$(Q,A) \sim \mathcal{T}(\widehat{I})$0

when $(Q,A) \sim \mathcal{T}(\widehat{I})$1. Third, if neither rule-based matching nor numeric parsing applies, an LLM-as-a-judge supplies binary reward (Wei et al., 12 Feb 2026). The training hyperparameters reported in the appendix are rollout group 8, epoch 1, AdamW with learning rate $(Q,A) \sim \mathcal{T}(\widehat{I})$2, weight decay $(Q,A) \sim \mathcal{T}(\widehat{I})$3, temperature 1.0, and gradient clipping at 1.0. The resulting models are referred to as ZwZ.

A central design claim is that the box overlay is not merely a metadata device. The paper argues that placing the box directly on the image supplies privileged information that compels the visual encoder and cross-modal projector to align the full-image representation with the micro-region (Wei et al., 12 Feb 2026). This suggests that the grounding transformation is the mechanism by which privileged crop information becomes available to a full-image model at test time without explicit zooming.

4. ZoomBench and the dual-view protocol

To evaluate whether zooming has in fact been internalized, the paper introduces ZoomBench, a hybrid-annotated benchmark of 845 VQA items spanning six fine-grained perceptual dimensions: fine-grained counting, OCR, color attributes, structural attributes, material attributes, and object identification (Wei et al., 12 Feb 2026). The benchmark is explicitly designed to quantify the difference between performance on the full image and performance on the corresponding key-region crop.

The construction process follows the same region-centric logic as R2I itself, but with human verification rather than human-authored question writing. In the appendix, benchmark synthesis uses Gemini-2.5-Pro as both question and answer generator. Three PhD-level paper authors independently verify samples while viewing the full image, the crop, and the generated QA pair. They check that the question is unambiguous and answerable, and that the answer is correct from both the cropped and full-image views. Each annotator spends about 10 hours verifying roughly 650 raw samples; 1,960 raw samples are checked in total, and after filtering overly easy items, 845 remain (Wei et al., 12 Feb 2026).

The final benchmark includes both open-ended and multiple-choice questions. The appendix specifies 224 open-ended questions with canonical target answers and 621 multiple-choice questions. The questions are described as diverse and non-templated. This hybrid format is intended to preserve both scalability and robustness to answer normalization issues.

ZoomBench further defines a dual-view protocol: Global-View uses the full image, while Regional-View uses the corresponding key-region crop. The performance difference between these two settings is termed the “zooming gap,” interpreted as the amount of evidence retrieval failure under realistic full-image conditions (Wei et al., 12 Feb 2026). The benchmark therefore measures not only task accuracy but also the extent to which a model can recover micro-evidence without explicit crop assistance.

5. Empirical performance and ablation evidence

A major empirical finding is that the zooming gap remains substantial even for very large MLLMs (Wei et al., 12 Feb 2026). Selected figures reported for ZoomBench are summarized below.

Model	Global / Regional	Gap
Qwen3-VL-8B	37.87 / 63.08	25.21
GLM-4.5V	— / —	24.38
Qwen3-VL-235B	— / —	22.96
GPT-5.1	— / —	22.84
Gemini-3-Flash	— / —	19.05
ZwZ-8B	58.11 / 73.37	15.26

Within this evaluation, ZwZ-8B achieves the smallest reported gap, at 15.26 points, while also improving Global-View performance to 58.11 (Wei et al., 12 Feb 2026). The paper further notes that counting is the hardest dimension, with regional access helping less than might be expected because counting requires both localization and aggregation. Structure and material exhibit the largest average gaps, whereas OCR and identification are comparatively easier.

On the main benchmark table, R2I training consistently improves general perception, specific perception, and out-of-distribution cognition. ZwZ-8B increases the average score of Qwen3-VL-8B from 61.52 to 68.12, with ZoomBench rising from 37.87 to 58.11. ZwZ-4B reaches 66.86 average, ZwZ-7B 62.42, and ZwZ-8B 68.12, which is reported as the best overall average among open-source models. The paper also states that ZwZ variants outperform larger open-source systems such as Qwen3-VL-235B, Kimi-K2.5, and GLM-4.5V on the averaged general-perception metric, and that they improve out-of-distribution benchmarks including MMStar and BabyVision. Additional gains are reported on AIGC detection and GUI-agent tasks such as LOKI, FakeCLUE, ScreenSpot Pro, and OSWorldG, where ZwZ-8B consistently surpasses the base Qwen3-VL-8B (Wei et al., 12 Feb 2026).

The comparison with TwI systems is central. On a dedicated agentic comparison table, ZwZ-8B obtains 81.9 average across VStar, HR-4K, HR-8K, and MME-RW-en, outperforming or matching many tool-using systems, including Pixel-Reasoner, DeepEyes, Thyme, DeepEyesV2, Mini-o3, SenseNova-MARS, and Skywork-R1V4. The paper also reports that ZwZ models are around $(Q,A) \sim \mathcal{T}(\widehat{I})$4 faster in inference speed than agentic or tool-use baselines while delivering higher average accuracy, because they operate with a single forward pass on the full image (Wei et al., 12 Feb 2026).

The ablation results support the crop-to-full-image design. Compared with “Direct Synthesis,” defined as global-to-global teacher–student distillation where the teacher sees the full image, R2I with box overlays yields markedly better performance. The grounding ablation on ZoomBench is especially explicit.

Variant	ZoomBench
Direct synthesis	40.95
No bbox	46.27
Bbox-in-question	46.98
Bbox-in-image	52.90

The reported ordering indicates that image-space box overlays are more effective than textual coordinate grounding or no grounding at all (Wei et al., 12 Feb 2026). The full 74K training set also outperforms a 10K subset, although the 10K subset already provides a strong boost. The paper further states that the distilled 74K set exceeds the performance of larger synthetic datasets such as Oasis, MM-Self-Instruct, DeepEyes data, Thyme-RL data, and TreeVGR-RL data, despite being smaller than Oasis’s 500K.

6. Interpretability, scope, and boundaries of distillation

The paper includes an interpretability analysis based on relative attention maps. For a key bounding box $(Q,A) \sim \mathcal{T}(\widehat{I})$5, attention coverage is defined as

$(Q,A) \sim \mathcal{T}(\widehat{I})$6

Using this measure on ZoomBench, ZwZ models allocate more attention mass to the annotated key region than their Qwen baselines: QwenVL-4B has 17.34% coverage versus 21.45% for ZwZ-4B; QwenVL-7B has 12.44% versus 13.39%; and QwenVL-8B has 17.39% versus 21.64% (Wei et al., 12 Feb 2026). This suggests that the distilled models more consistently concentrate question-relevant attention within the micro-region.

The broader claim is deliberately bounded. The paper does not argue that all zooming or tool use is dispensable. Instead, it distinguishes information-neutral actions from information-gain actions. Zooming, cropping, flipping, rotating, denoising, 2D or 3D grounding, and drawing auxiliary structures are presented as information-neutral: they do not introduce content unavailable in the image, but rather reorganize access to existing evidence. Web search and external retrieval, by contrast, are information-gain actions and therefore are not distillable in the same way because they reveal genuinely new content (Wei et al., 12 Feb 2026).

This boundary conditions the paper’s central thesis. A plausible implication is that R2I is most effective when the bottleneck is evidence access rather than the absence of external world knowledge. The reported results support that interpretation by showing substantial reductions in the global–regional zooming gap and strong single-forward-pass performance. At the same time, the authors identify limitations: the current data pipeline has not broadly incorporated spatial reasoning and multi-object perception tasks, and TreeBench was not evaluated because those abilities were not central to the present setup (Wei et al., 12 Feb 2026). Future extensions are suggested in the direction of spatial reasoning tools and multi-object search tools.

Taken together, Region-to-Image Distillation defines a specific paradigm for compressing the benefit of agentic zooming into model parameters. Its central insight is that, for many fine-grained perceptual failures, the decisive advantage of zooming lies not in the act of interactive tool use itself, but in improved access to already present visual evidence; when that access can be recreated during data generation and grounded back onto the full image, much of the benefit can be retained at test time without repeated tool calls (Wei et al., 12 Feb 2026).