Images

The Visual Token Audit

A vision model reads a picture as image tokens, and a scanned page costs thousands of them. We measure what your images cost in your model’s own tokens, and how many of those tokens a visual lexicon removes at a fidelity we state in pixels — before anything in your stack changes.

2.60×

Fewer image tokens on scanned documents, with no measurable change in answers on two open vision models — 200 questions each

3.07×

Fewer image tokens on rendered text pages, with a pixel-identical round trip and no measurable change in what the model reads back

4.27×

On scanned documents when large pages are capped at a standard resolution first — still no measurable change, on the smaller model

No model

The audit itself runs on CPU and loads no model, so measuring your images needs no GPU and no cloud

What the audit establishes

Three figures, each specific to your images and each reconstructible from the measurement that produced it.

What your images cost today

Image tokens per image on your model’s own grid, at the resolution limit you run, with the spread across your images rather than an average alone.

What each fidelity tier removes

The reduction at a pixel-identical round trip, and at stated near-lossless tolerances — each with the largest pixel error it actually incurred and its PSNR.

Where it concentrates

The reduction by image size and per image, so the work starts where your images carry the most redundant area — typically documents, forms and rendered pages.

Measured on public benchmarks

Training-free, on two open vision models. The same questions, answered with the full image and with the reduction.

Fidelity is stated per row: pixel-identical, or every pixel within 16 or 32 of 255 levels of the original. On every row, answers showed no measurable loss against the full image, on either model.
Images	Fidelity	Fewer image tokens
Scanned documents, questions answered (DocVQA, 200)	within 32 of 255	2.60×
Rendered text pages, read back (20 pages)	pixel-identical	3.07×
Receipts, read back word by word (CORD, 100)	within 32 of 255	2.36×
Receipts, totals extracted (CORD, 193)	within 16 of 255	1.43×
Text in photographs (TextVQA, 200)	within 32 of 255	1.37×
Mixed reading tasks (OCRBench, 200)	within 32 of 255	1.76×

Every figure is paired: the same questions, answered with the full image and with the reduction, on the same images. On receipts read back word by word, scores went up rather than down. Each reduction was also measured against simply shrinking the image, and on rendered pages and receipts read word by word the lexicon came out ahead, even where the shrunken image kept more tokens.

A page can cost less as a picture than as text. On rendered text pages, the larger of the two models read a page from 251 image tokens on average and got 98.98% of its characters right; the same page as text averages 461 tokens in that model’s own tokenizer: 1.84× fewer as a picture.

How the engagement runs

1 — ImagesYou provide a representative folder of the images your models process: documents, forms, receipts, pages. Where data handling is constrained, the transfer is scoped to your requirements before anything moves.

2 — MeasureImage tokens per image on your model’s grid at your resolution limit, and what each fidelity tier removes, with the pixel error each one actually incurs.

3 — DeliverThe three figures above, per image and in aggregate.

4 — Optionally, your answersFor open models we can run, the same questions answered with the full image and with the reduction, on your tasks, with the paired change reported — before anything is committed.

Scope, timeline and fee are set against the number and kind of images. Tell us what you run and we will come back with both.

Next step

Tell us what your models look at

The kind of images, the model you run, and roughly how many a day. That is enough for us to scope the audit.

Request a visual audit →