Straight answers to the questions we hear most often about Image to PNG Converter. Everything here reflects how the plugin actually behaves in version 1.3.0; if something on this page and the plugin disagree, the plugin is right and we would like to know.
The short version. Your images never leave your device. Rectangle Crop and Freehand Lasso make no network requests at all. The two AI modes download a model to your browser the first time they are used, and that is the only thing ever fetched from outside your site.
Privacy and data
Are my images uploaded anywhere?
No. Decoding, selecting and PNG encoding all happen in the visitor’s browser using the Canvas API. The image is never sent to the site’s server or to any third party, and the plugin stores nothing in the WordPress database.
This is true of the two AI modes as well. What is downloaded is the model; the picture itself is processed on the device and stays there.
What is downloaded when someone uses one of the AI modes?
Rectangle Crop and Freehand Lasso are fully offline. The two AI modes each fetch a model the first time they are used in a session:
- Auto Subject Extraction downloads roughly 44 MB of model data (with the default
isnet_quint8model) from staticimgly.com, plus the @imgly/background-removal library from jsDelivr. - AI Click Select downloads about 38 MB of model files (a 22.2 MB vision encoder and a 15.8 MB prompt encoder and mask decoder) from the Hugging Face CDN, plus roughly 4 MB of ONNX Runtime WebAssembly from jsDelivr.
The browser caches these files, so later images in the same session are much quicker. Transformers.js also keeps the click-select model in the browser’s Cache Storage, so a returning visitor does not pay for the download again. On a phone on mobile data, that first download is still significant, which is why both modes are optional.
Do I need to mention the plugin in my privacy policy?
For site owners
The plugin itself sends nothing anywhere, stores nothing on your server and adds nothing to the database. If you keep the AI modes enabled, though, a visitor’s browser will contact third-party hosts to fetch the models: cdn.jsdelivr.net and staticimgly.com for Auto Subject Extraction, and cdn.jsdelivr.net and huggingface.co (including cdn-lfs*.huggingface.co) for AI Click Select. Those files are then kept in the visitor’s browser cache.
If you publish a privacy policy or a cookie notice, you may need to disclose those hosts. This is not legal advice; please check the requirements that apply to [Your Company Name] in [your jurisdiction].
Can I make the tool completely offline?
For site owners
Yes. Turn off both AI modes in the shortcode:
Convert your images to PNG right inside your browser. Nothing is ever uploaded to a server. Convert Any Image to PNG
Rectangle Crop and Freehand Lasso remain available, both fully offline, and nothing external is ever requested. You can also switch off just one of the two: ai="no" removes Auto Subject Extraction and sam="no" removes AI Click Select.
Using the tool
Which image formats are accepted, and how large can a file be?
The tool accepts JPG, JPEG, PNG, WEBP, GIF and BMP. You can drag and drop a file, click to browse, or paste an image straight from the clipboard.
By default a file may be up to 20 MB. The site owner can set a different limit with the max_size attribute, anywhere from 1 to 100 megabytes; values outside that range fall back to the default of 20. A file over the limit is declined with a clear message rather than failing silently.
How do I convert a whole image without cropping it?
When an image opens in the editor, the selection box already covers the whole image. Simply press Confirm and Convert to PNG without moving the box. If you have already moved or resized it, Reset Selection puts it back over the full image.
The result appears with a preview and a Download PNG button, and the file keeps the original filename with a .png extension. You can convert as many images as you like in one session without reloading the page.
Why is my converted PNG larger than the original JPG?
That is expected. PNG is a lossless format, so re-encoding a lossy JPG or WEBP usually produces a bigger file. The upside is that no further quality is lost: the PNG holds exactly the pixels that were in your image, and it will not degrade if you edit and save it again.
If file size matters more than lossless quality for a particular image, keep the original alongside the PNG.
Is the crop exact, even on a phone?
Yes. The selection is stored in the source image’s own pixel coordinates, not in the on-screen size, so a 200 x 100 selection always produces a 200 x 100 PNG taken from exactly that region, whether the image is displayed at full size on a desktop or scaled down on a phone. The readout inside the selection box shows the selected width and height live, in the image’s real pixels.
The same goes for the other modes. Freehand Lasso exports the inside of your outline trimmed to its bounding box, and AI Click Select sets transparency one pixel at a time from a full-resolution mask, so the cut-out is exact rather than approximate.
What happens to animated GIFs?
A PNG file holds a single still image, so an animated GIF is converted using its first frame only. The other frames are not included. If you need a different frame, save that frame as a still image first and convert it on its own.
How long do the AI modes take, and what is the difference between WebGPU and WebAssembly?
It depends on the device and the connection. On a measured desktop run, a first Auto Subject Extraction took about 18 seconds including the model download, and a repeat about 10 seconds.
AI Click Select is split into two steps. Entering the mode runs the image encoder once, which is the slow part; you will see “Analyzing image with AI…” while it works. Every click after that runs only a small decoder against the cached result, so the first outline takes seconds and later clicks take a fraction of a second. Switching to another mode and back does not re-encode the image. Measured on a desktop using WebAssembly, the model and runtime loaded in about 4 seconds, the encoder took about 9 seconds and each click 0.3 to 0.5 seconds.
WebGPU and WebAssembly are the two ways a browser can run the model. WebGPU uses the device’s graphics hardware and is considerably faster; WebAssembly runs on the main processor and is supported by more browsers. With the default sam_device="auto" setting, AI Click Select uses WebGPU when the browser has it and falls back to WebAssembly otherwise, so there is nothing for the visitor to choose. A site owner can force one or the other with sam_device="webgpu" or sam_device="wasm". Low-powered phones will be slower than the figures above.
Troubleshooting
It says “AI selection isn’t available right now. Try Rectangle Crop or Freehand Lasso instead.” What does that mean?
AI Click Select shows this message when its library or model could not be fetched, or when the browser cannot run it. Auto Subject Extraction shows a similar message, pointing you at the rectangle and freehand modes, when its library cannot be downloaded (no connection, blocked CDN) or the browser cannot run it. In both cases Rectangle Crop and Freehand Lasso keep working, so you can still finish the job by hand.
The usual causes are:
- You are offline or on a very poor connection, and the model download could not complete.
- A content blocker, corporate firewall or strict Content-Security-Policy is blocking cdn.jsdelivr.net, huggingface.co or staticimgly.com.
- The browser is too old to run the model.
If you are a visitor: check your connection and try again, try the page with any content blocker paused, or use a current browser. Pressing Reset Selection discards the failed result and lets you try the mode again.
If you own the site: check that your Content-Security-Policy allows the hosts listed in the next answer, and that any ai_url or sam_url override is an https URL that actually serves the library.
Which hosts must my Content-Security-Policy allow for the AI modes?
For site owners
If you run a strict Content-Security-Policy, the AI modes need these hosts allowed. The manual modes need nothing beyond your own site.
| Mode | Host | What it serves |
|---|---|---|
| Auto Subject Extraction | cdn.jsdelivr.net | The @imgly/background-removal library |
| Auto Subject Extraction | staticimgly.com | The model and WebAssembly files |
| AI Click Select | cdn.jsdelivr.net | Transformers.js and the ONNX Runtime WebAssembly files |
| AI Click Select | huggingface.co and cdn-lfs*.huggingface.co | The model weights |
If you point ai_url or sam_url at a different host, allow that host instead. If you disable both AI modes with ai="no" sam="no", no extra hosts are needed at all.
The tool says my image is too large. What are the limits?
There are three separate checks, and the message tells you which one applied:
- File size. The default limit is 20 MB per file. The site owner can raise or lower it with the
max_sizeattribute (1 to 100 megabytes). - Pixel dimensions. Images with over-large dimensions are declined with a clear message. Since version 1.1.0 this check applies to the selected area rather than the whole image, so try Rectangle Crop and select a smaller region of the photo.
- AI Click Select only. Images above roughly 16 megapixels are declined in this mode, because the model returns full-resolution masks and several of those at once would be unreasonable in a browser tab. Use one of the manual modes for those images, as the message suggests, or reduce the image’s dimensions before choosing it.
The tool says my browser is not supported. Which browsers work?
Any current version of Chrome, Firefox, Safari and Edge. Older browsers that cannot encode canvas content are shown a clear message instead of the tool silently failing, and the fix is to update the browser.
The AI modes have their own requirements on top of that. If the browser cannot run the model, the card shows the message described above and the two manual modes keep working. WebGPU is used for AI Click Select only where the browser supports it; everywhere else the tool falls back to WebAssembly automatically.
I changed the ai_url attribute and now see “Failed to resolve module specifier”.
For site owners
An ai_url override must be an https URL to a dependency-resolved ESM build. The package’s own dist/index.mjs imports its onnxruntime-web peer dependency by bare specifier, which a browser cannot resolve: the module loads successfully and then fails at inference time with exactly that message.
jsDelivr’s /+esm endpoint rewrites those imports, which is why the default is https://cdn.jsdelivr.net/npm/@imgly/background-removal@1.7.0/+esm. Unless you have a specific reason to change it, leave ai_url at its default.
The same caution applies to sam_url. The default pins Transformers.js 3.5.0, the version the plugin was tested against. Newer releases exist; if you point sam_url at one, re-test before doing so on a live site.
The tool appears but does nothing when my theme loads it over AJAX.
For site owners
The plugin’s stylesheet and script are enqueued only on pages where the shortcode actually renders, so nothing is added to the rest of your site. If your theme or page builder injects the shortcode after the page has loaded, the script has no chance to wire up the new markup. Call window.ITPNGC.init(); once the markup is in the DOM and the tool will come to life.
If you want the tool in a template file rather than the editor, use Convert your images to PNG right inside your browser. Nothing is ever uploaded to a server.do_shortcode( ' in PHP rather than pasting raw HTML, so the assets are enqueued correctly. In the editor, a Shortcode block, a Custom HTML block or the Classic editor all work.Convert Any Image to PNG
Still stuck? Email [contact@example.com] and include the browser name and version, the exact message shown on the card, and the shortcode you used (with any attributes). That is usually enough for us to reproduce the problem.