Home / Annotation & accounting / AI annotation

AI annotation

Upload 1 to 20 images, describe the target in one sentence (e.g. "mark the red chillies") and choose whether boxes should be axis-aligned or at any angle; the AI inspects each image and returns the four corner coordinates and four edge lengths of the clearest target it finds — one box per image — ready to copy as text or JSON. Exporting can draw the box if you want it, and without that option it is the original image. Charged per image (¥0.5 each) via Alipay, detection starts after payment. Charging is per image and so are refunds: when only some images fail, the fee for the failed ones is refunded automatically while the results for the rest are delivered as usual; when the whole batch fails the entire order is refunded, and if the model call succeeded but produced no usable boxes, the fee is not refunded.

Pay per run (¥0.5 per image), no sign-up, Alipay online payment only. The original image is sent to our backend for AI inference only after payment succeeds; nothing is written to server storage, selection and coordinate mapping happen in your browser, and failed runs are refunded automatically.

More tools in Annotation & accounting

Frequently asked questions about AI annotation

How is AI annotation different from manual annotation?
With manual annotation you drag the box yourself, so the coordinates are exactly where you drew them, at no cost and with no upload. AI annotation hands the images to a large model that returns one box per image, around the clearest match for your prompt — add any other targets in the same picture with manual annotation. It suits large batches or when you would rather not draw boxes one by one. Corner order, edge mapping and export file names are identical across both pages, and you can always fine-tune AI results on the manual page.
How is it priced? Is there a discount for many images?
Priced per image: ¥0.5 each, so five images cost five times that, with at most 20 images per order and no discounts or extra fees. Before payment the page spells out "unit price x count = total". Charging is per image and so are refunds: when only some images fail, the fee for the failed ones is refunded automatically while the coordinates and edge lengths of the remaining images are still returned; when the whole batch fails, the entire order is refunded. If the model call succeeded but produced no usable results (typically because the prompt does not match the picture), the fee is not refunded.
Some images failed — can I get a partial refund?
Yes — charging is per image and so are refunds: when recognition finishes, the fee for the failed images is refunded automatically (fail 2 out of 5 and you get 2 images’ worth back) while the coordinates and edge lengths of the remaining images are still returned, with the refund number shown in the results area. Note the policy: images where the model call succeeded but produced no usable annotation targets (typically because the prompt does not match the picture) are not refunded — the fee has been incurred; when the whole batch fails (call error or error-code response), the entire order is refunded automatically. So before paying, describe the target clearly (e.g. "mark the red chillies", "person wearing a yellow hard hat") and pick sharp images where the target is not hidden; after detection you can check the preview and the coordinates image by image, and re-run the images you are unhappy with as a separate order using a different prompt.
Are the coordinates in source-image pixels? How accurate are they?
Yes, they are source-image pixels: the model only returns normalised 0 to 1 positions, the backend converts them with the real width and height it read, and the four corners and edge lengths are computed with the very same formulas as manual annotation, so the numbers match what you would draw by hand. Accuracy depends on the picture — clean, unobstructed, sparse subjects are usually spot on, while tiny, overlapping or occluded subjects can drift, so review each image before using the results.
Do exported images include the annotation line?
Not by default: you get the original image itself (pixel for pixel), with the coordinates and edge lengths carried in the file name, so it works straight away as "image plus coordinate labels". Tick "Draw the annotation line on the exported image" on the left when you want to see the box, and optionally number the corners 1-4. Both ways produce the same file name format.
How many images can I upload, and how large may they be?
One to 20 images per order, at most 15MB each, PNG or JPEG only (longest side up to 4096 px, aspect ratio up to 3:1 — compress or crop larger ones first). Images are uploaded to our site (deleted automatically after 24 hours) and then passed to the model. The prompt only needs to name the target, e.g. "red chilli", "yellow hard hat" or "QR code".
What is the difference between a horizontal box and any angle?
A "Horizontal box" keeps all four edges parallel to the image, which suits upright subjects such as documents, cards and tables; "Any angle" adds a round handle once the box is drawn - drag it to rotate the box to any angle (hold Shift to snap every 15 degrees) or type the angle on the right, which suits tilted subjects. Both kinds produce coordinates and edge lengths by exactly the same rules, so you can switch at any time.
Does AI annotation draw the boxes for me?
Yes. In "AI annotation" you upload 1 to 20 images and describe the target (e.g. "mark the red chillies"); the model finds the objects and returns their four corner coordinates, so you never drag a box yourself. Results can be copied per image or exported as annotated images, and anything you dislike can be fine-tuned on the manual annotation page.
Are the coordinates relative to the original image or to the size shown on screen?
They are real pixels of the original image. The image is scaled down to fit the page, but both the box and the copied coordinates are recorded in source-image pixels, so the same spot gives the same numbers on a different screen or browser.

About Local Toolbox

Local Toolbox gathers the small tools you reach for every day into one place: formatters, encoding converters, date maths, ID card and address parsing, image and QR code handling, password and amount conversion. Open a page and start working — no sign-up, and almost every tool has no usage limit.

Most features compute inside your browser, so your input never leaves your device; the pages are plain static assets, load fast and keep working offline. The IP page asks our backend for your outbound address when it opens and sends none of your input. AI watermark removal, AI background removal, AI image generation and AI upscaling send content to our backend for inference (watermark removal is ¥1 per run; background removal is ¥1 per run; image generation and upscaling are priced by resolution — 2K ¥2, 4K ¥3); AI annotation sends images to a large model that locates the target and returns coordinates (¥0.5 per image, refunded only when every image in the batch fails); AI flowchart, AI mind-map and AI Mermaid chart generation send your text description to a model that builds the diagram (three separate products, charged ¥0.5, ¥0.5 and ¥0.5 per run respectively); and the accounting helper sends the income, expense, social-insurance and balance figures you fill in to generate a report (¥0.5 for the spoken analysis, ¥1 for the report). These features accept Alipay only, upload nothing before payment succeeds, and refund automatically if processing fails.

Links