Image to LLM Optimizer
Drop a heavy image here
4K screenshot, photo, or scanned doc — drag & drop, or click to browse
100% private — stays in your browser, never uploaded
JPG, PNG, WebP · up to 15MB
Optimize
Sample result — a 1200×800 invoice photo, 1.4MB, at Extract Text. Drop your own image to replace it.
OCR the image to plain text — best for invoices, documents, and code screenshots. Most accurate mode — reads the actual words, no guessing.
Before
1,105
tokens (high)
After
149
tokens
Saved
87%
$0.00478 · ₹0.3991
100% accurate — these are the actual words, not a model's guess at them.
Sample extraction — a real invoice photo. Drop your own image and click Extract to replace it.
149 text tokens vs 1,105 vision tokens — ~4 chars/token estimate
Resizing, cropping, and OCR all run in this tab — your image never leaves the device.
How GPT-4o vision tokens work, and how to save 90%
GPT-4o and every model built the same way don't look at an image's file size — they look at its pixel dimensions and split it into 512×512 tiles. In low detail mode, the image is resized to fit inside 512×512 before the model ever sees it, so it costs a flat 85 tokens whether the original was 100KB or 10MB. In high detail mode, the model tiles the image at its real resolution — 170 tokens per tile plus an 85-token base — so a 1024×1024 image (4 tiles) costs 765 tokens, and a 4000×3000 phone screenshot (48 tiles) costs 8,245 tokens, over $0.04 on GPT-4o's own per-token pricing, before you've asked it anything.
This is why compressing an image's quality — the TinyPNG/Compressor.io move — saves disk space and saves zero vision tokens: quality compression never touches pixel dimensions, and tiling only ever looks at pixel dimensions. Resizing the image is what saves tokens, because it changes how many tiles fit across it.
Three ways to cut the bill
- Smart Resize4000×3000 → 1024×1024 cuts 48 tiles to 4 — about 90% fewer tokens, same visual understanding for most questions.
- Extract TextOCR turns a screenshot or invoice into plain text billed as ordinary text tokens — no tiling at all, often a 95%+ cut.
- Smart CropSend only the part of the frame that answers the question — a smaller image tiles into fewer tiles.
When to use which
If the model just needs to recognize what's in the picture — a product photo, a meme, "is there a person in this frame?" — Low detail's flat 85 tokens is almost always enough. If it needs to read fine print — a receipt, a chat screenshot, a code editor — High detail at a resized 1024px width keeps text legible for a fraction of the raw-resolution cost. If the image is the text — a scanned invoice, a document photo — OCR beats both: the model reasons over the extracted words directly, at plain text-token pricing, with none of the tiling formula involved at all.
Keep the original when the visual detail is the point — a chart's exact shape, a photo's composition, a diagram's layout. This tool resizes, crops, and OCRs; it doesn't change what a model can see, only how much it's billed to look.
For text-heavy documents rather than images, Toolsda's PDF to LLM Optimizer does the same job for PDFs — removing repeated headers, converting tables to Markdown, and keeping only the pages that matter. For pure file-size compression (not token reduction), see the Image Compressor.
Related tools
- Token Counter — count tokens and cost for text you already have, across nine models.
- PDF to LLM Optimizer — the same token-saving idea, for PDFs instead of images.
- Image Compressor — shrink a file to an exact KB target; use this tool instead when the goal is fewer vision tokens, not a smaller download.
Frequently asked questions
How does GPT-4o count image tokens?
By tiles, not kilobytes. At "high" detail, GPT-4o splits an image into 512×512 tiles, charges 170 tokens per tile plus an 85-token base, and bills that regardless of the file's compression or format. A 4000×3000 screenshot needs ceil(4000/512)×ceil(3000/512) = 8×6 = 48 tiles, so it costs 85 + 48×170 = 8,245 tokens — about $0.04 on GPT-4o input pricing, before you have asked a single question about it.
What is the difference between low and high detail mode?
Low detail resizes any image to fit inside 512×512 internally and charges a flat 85 tokens, whatever the original resolution — good for "what is in this image?" but too coarse to read small text. High detail tiles the image at full resolution so the model can read fine print, at the 170-tokens-per-tile cost above. Most vision API calls default to high detail unless you explicitly set `detail: "low"`.
Does compressing an image's quality save tokens?
No — and this is the mistake nearly everyone makes. TinyPNG-style quality compression shrinks the file in kilobytes; it does not change the pixel dimensions, and tiling only ever looks at pixel dimensions. Taking a 4000×3000 photo from quality 100% to quality 20% saves disk space and saves exactly zero vision tokens. Resizing it to 1024px wide is what saves tokens, because fewer tiles fit across a smaller image.
How do I save 100% of the vision tokens on an image?
Extract the text instead of sending the picture. OCR (Extract Text mode here) turns a screenshot, invoice, or document photo into plain text billed at ordinary text-token rates — typically a few hundred tokens instead of thousands — with no tiling at all. This only works when the content is genuinely text; a photo, chart, or diagram still needs to be sent as an image for a model to reason about it visually.
OCR vs sending the image — which one actually saves more tokens?
It depends on how much of the image is text. A 1200×800 invoice photo costs 1,105 vision tokens at high detail; OCR reads the same words in roughly 150 text tokens — an 86% cut, and the text is 100% accurate rather than a model's visual guess at small print. This tool's auto-detect banner and the Cost Breakdown tab's comparison row do this math for your actual image: once the extracted text comes in comfortably under the image's token count, OCR is recommended. A photo, chart, or diagram with little or no text won't benefit — there's nothing for OCR to extract, so Smart Resize is the better mode there.
Is my image private with this tool?
Yes. Every operation — reading the file, resizing, cropping, and OCR — runs in your own browser tab using Canvas and, for text extraction, Tesseract.js compiled to WebAssembly. Nothing is uploaded to Toolsda's servers or anyone else's; you can confirm this by opening your browser's Network tab while you use the tool, or by disconnecting from the internet after the page loads.
How much does a 4K image actually cost in GPT-4o?
A 4000×3000 image at high detail is 48 tiles — 8,245 tokens, or roughly $0.04 of input cost on GPT-4o's per-token pricing for that single image, before any text in the conversation. Resize it to 1024px wide first and the same picture drops to 4 tiles — 765 tokens, about 91% cheaper — and a model answering "what does this screenshot show?" rarely needs more resolution than that.