Know When to Zoom: How Small Is Too Small for Vision-Language Models?

Vohra, Siddharth; Xu, Min

Know When to Zoom

Online services shrink a large image, or reject it, before the model ever sees it. Small details go first.

We measured how much detail survives for seven vision-language models from four vendors, with an eye exam of printed codes and 601 questions about small objects in real photos. Zooming wins much of the detail back, and Claude’s own reply tells an agent when it is needed.

Read the paper Code and data

One car, 20 pixels wide, in an 8000 × 6000 drone photo

In the photo

20 pixels wide

What Claude receives

About 6 pixels wide

OpenAI’s default

Over 30,000 patches of 32 by 32 pixels

Rejects the photo outright

One square per pixel. The car’s shape is illustrative; its widths are the paper’s.

The paper · SENSE Workshop · WACV 2027

Know When to Zoom: How Small Is Too Small for Vision-Language Models?

Siddharth Vohra, Min Xu

Accepted to the SENSE workshop at WACV 2027, to be presented in January 2027 in Orlando, Florida.

Online services shrink large images before a vision-language model sees them. An eye exam and 601 questions about small objects across seven models show how much detail survives, how much zooming wins back, and when an agent should zoom.

Abstract

Online vision-language models often miss small details in large images, because the service shrinks the image, or rejects it, before the model ever sees it. A car 20 pixels wide in an 8000 by 6000 drone photo reaches Claude about 6 pixels wide, and OpenAI’s default rejects the photo outright. We measure how much detail survives for seven models from four vendors, using an eye exam of printed codes and 601 questions about small objects in real photos. The smallest text a model reads half the time grows from 6 pixels or less on a small image to a predicted 10 to 13 pixels on a 7680 by 4320 photo at default settings. Zooming wins much of the detail back. Splitting each photo into up to nine tiles raises accuracy by 11.1 to 14.5 percentage points for Claude and Gemini, and by avoiding rejections it lifts GPT-6.1 Sol from 50.7% to 83.7%. A two-step zoom, in which the model first names the region it needs, matches or beats tiles on Gemini and comes within 2.5 points of them on Claude, with about 40% of their image tokens. When a detail is lost, Claude and Gemini both say they cannot tell about half the time, but they fail differently in the rest: Gemini gives a confident wrong answer in 43% to 48% of these cases, Claude in only 5% to 9%. A check based on detail size, run before the call, misses most of the gain from tiles on Claude and Gemini, but Claude’s own reply is a reliable trigger: tiling only after an unsure reply comes within 1.4 points of always tiling with 62% to 70% of the image tokens. We turn these findings into five rules for when an agent should zoom, and release the eye exam, every model answer and code at https://siddvoh.com/know-when-to-zoom.

This page is a plain-language walkthrough. The paper is the source to cite, and its reference and BibTeX are at the end of the page.

What happens before the model looks

Many questions about images turn on small details. A satellite scene may show a vehicle only a few pixels wide, and a security camera frame may hold a license plate in one corner.

Every service we tested fits images to its own limits before the model sees them. A user who does not know these rules cannot tell whether a wrong answer came from the model or from the shrinking.

What each service does to a large image

From each vendor’s documentation, checked against the token counts the services reported.

Claude: Shrinks the image to at most 2576 pixels on its long side and 4,784 tokens, each a patch of 28 by 28 pixels.
Gemini: Turns any image into at most 280 to 2,240 tokens, depending on a resolution setting. Leaving it unset gives the high setting.
OpenAI: The default keeps the full image up to 30,000 patches of 32 by 32 pixels and rejects anything larger. The high setting shrinks it to at most 2,500 patches.
Qwen: Its host keeps up to 16,384 tokens of 32 by 32 pixels, far more pixels than the others.

models: two each from Anthropic, Google and OpenAI, and the open model Qwen.

601

multiple-choice questions about small objects, from V*Bench, HR-Bench 8K and MME-RealWorld-Lite.

canvas sizes in the eye exam, from 1024 by 768 to 8000 by 6000 pixels.

An eye exam for models

Much like the letter chart at an optometrist’s office, each test image shows a random five-character code, printed at one of ten text sizes on a plain gray background. We call the smallest text size a model reads half the time its readable size, and print the same sizes on four canvases, from small to large. The larger the canvas, the more a service shrinks it.

Smallest cap height read half the time, at each model’s default setting

Canvas 1024 × 768 2048 × 1536 4096 × 3072 8000 × 6000

Haiku native 4.7 px 4.8 px 8.1 px 14.4 px

Sonnet native 4.2 px 4.4 px 6.5 px 12.0 px

Flash-Lite high 4.8 px 4.9 px 8.2 px 13.7 px

Flash high 4.2 px 4.4 px 7.5 px 13.2 px

Luna auto 4 to 6 px 4.2 px 4 to 6 px rejected

Sol auto 4.3 px 4.5 px 4 to 6 px rejected

Qwen default 5.4 px 6.2 px 9.0 px 15.1 px

Each letter is drawn at three times the cap height the model reads half the time. Where the fit does not cross half within our sizes, the cell gives the sizes between which reads did. OpenAI’s default rejects the largest canvas.

On the smallest canvas, every model’s readable size was 6 pixels or less. Most settings shrink the two largest canvases, and there the readable size grew. Sonnet needs 4.2 pixels on the 1024 canvas but 12.0 on the 8000 canvas, and Flash at high, which matches its default, needs 4.2 and 13.2.

For Claude, the documented resize explains much of what the model can read. It fits Haiku closely and Sonnet less well. Gemini sees more than its token count: at high, Flash could read letters 0.15 of a token tall on the 1024 canvas, but letters only 0.06 of a token tall on the 8000 canvas.

What one 8K photo keeps

Text a model reads half the time on a 7680 × 4320 photo

Predicted from the eye exam at default settings. Smaller text is read less often.

Sonnet: Letters 11 pixels tall.
Haiku: Letters 13 pixels tall.
Flash and Flash-Lite: Letters 13 pixels tall at high, which matches the default.
Luna and Sol: The default rejects such a photo. The high setting would need 14 to 16 pixels.
Qwen: The fit gives 10 pixels, but this estimate is loose, since its readable size grew even on canvases its host keeps whole.

Zooming wins much of it back

We sent each question in up to five ways: as is, at the highest setting, as 4 tiles, as 9 tiles, or as a two-step zoom, in which the model first names the region it needs and then sees that region, cut from the original, next to an overview of the photo.

Accuracy on 601 questions about small objects in large photos

Photo as is 3 × 3 tiles Two-step zoom

Haiku as is 59.2% tiles 73.7% zoom 73.5%

Sonnet as is 72.4% tiles 83.5% zoom 81.0%

Flash-Lite as is 58.6% tiles 70.5% zoom 72.2%

Flash as is 68.1% tiles 80.5% zoom 83.0%

Luna as is 38.6% tiles 65.7% zoom 70.2%

Sol as is 50.7% tiles 83.7% zoom 82.2%

Tiles fall back to 2 × 2 for photos under 4000 pixels. The zoom uses about 40% of the tiles’ image tokens on Gemini. Qwen received photos only as is, and scored 71.5%.

Up to nine tiles raise accuracy by 11.1 to 14.5 percentage points for Claude and Gemini, and by avoiding rejections they lift GPT-6.1 Sol from 50.7% to 83.7%. The two-step zoom matches or beats tiles on Gemini and comes within 2.5 points of them on Claude, with about 40% of their image tokens.

Tiles helped even on V*Bench, whose median target is 36 pixels across its shorter side. On remote sensing, the two-step zoom scored highest for five of the six models, lifting Flash from 61.5% to 79.2% and Sol from 36.2% to 84.6%.

When the detail is lost

When the full photo misses a detail that tiles find, Claude and Gemini both say they cannot tell about half the time, but they fail differently in the rest. Gemini gives a confident wrong answer in 43% to 48% of these cases, Claude in only 5% to 9%.

Questions tiles got right and the photo as is got wrong. Each square is one question.

Said it cannot tell Wrong, confidence below 70 Wrong, confidence 70 or more Photo rejected

Haiku 98 questions, 9 confidently wrong

Sonnet 75 questions, 4 confidently wrong

Flash-Lite 109 questions, 52 confidently wrong

Flash 95 questions, 41 confidently wrong

Luna 194 questions, 17 confidently wrong, 156 rejected

Sol 210 questions, 12 confidently wrong, 195 rejected

Knowing when to zoom

A check based on detail size, run before the call, misses most of the gain from tiles on Claude and Gemini.

Claude’s own reply is a reliable trigger. On the lost-detail questions, Claude said it could not tell, or gave a confidence under 70, in 91% of the cases for Haiku and 95% for Sonnet, against only 52% for Flash-Lite and 57% for Flash. Tiling only after such a reply comes within 1.4 points of always tiling with 62% to 70% of the image tokens.

On Claude, tile when the reply is unsure

Send the photo as is and tile it whenever the reply says it cannot tell or gives a confidence under 70. Where accuracy matters most, tile every photo.

On Gemini, always use the two-step zoom

Its confident replies hide lost details. Check that each box is valid, and send the whole photo if it is not.

On OpenAI’s default, watch the patch limit

Use tiles for any image above 30,000 patches, since the service rejects it, or a zoom whose crop stays under that limit.

Check before the call

Use the service’s rules to see whether it will shrink the image. A size check like ours should use a detail size well below 16 pixels.

Doubt confident answers

Except on Claude, whose confidence warned of lost details in our tests, treat a confident answer about a small detail with suspicion.

The paper

SENSE Workshop · WACV 2027

Know When to Zoom: How Small Is Too Small for Vision-Language Models?

Siddharth Vohra, Min Xu

Sensing via Agentic Reasoning Systems (SENSE) Workshop at WACV 2027 · Orlando, Florida, USA · January 2027

Accepted to the SENSE workshop at WACV 2027, to be presented in January 2027 in Orlando, Florida.

The service rules for every setting, the eye-exam fits and tests, all ways of sending a photo for every model, and the zoom check.

Read the paper

If this is useful in your work, feel free to cite it as: Vohra, S. and Xu, M. (2027). Know When to Zoom: How Small Is Too Small for Vision-Language Models? Sensing via Agentic Reasoning Systems (SENSE) Workshop at WACV 2027 (to appear).

Cite this

.bib

@inproceedings{vohra2027zoom,
  title     = {Know When to Zoom: How Small Is Too Small for Vision-Language
               Models?},
  author    = {Vohra, Siddharth and Xu, Min},
  booktitle = {Sensing via Agentic Reasoning Systems (SENSE) Workshop at
               WACV 2027},
  year      = {2027},
  address   = {Orlando, Florida, USA},
  note      = {To appear},
  url       = {https://siddvoh.com/know-when-to-zoom/}
}

Code and data on GitHub The eye exam, every model answer, and the code that rebuilds every result above.

Every result on this page is in the paper and rebuilds from the released answers. The car in the drawing at the top is illustrative; its widths are the paper’s. Work by Siddharth Vohra here is unrelated to his position at Amazon Web Services.