Troubleshooting

Most problems with the background remover come down to three things: the one-time model download, which model your hardware gets, and the edges of difficult subjects. Find your symptom below. Every error message the tool can show is quoted exactly, so you can search this page for the words on your screen.

Why is the first photo so slow?

The model is downloading. It is about 98 MB on the GPU path and 46 MB on the CPU path, and it downloads once. On a typical connection that is 10 to 30 seconds; after that it is in the browser cache and the next photo starts straight away.

The page normally starts the download in the background as soon as it opens. It holds off when your browser has Data Saver on or reports a 2G or 3G connection, and then the download starts when you add a photo instead. In a private or incognito window the browser may refuse to cache the file, so it downloads again on every visit. That is slower but still works.

Why is every photo slow, even after the download?

Check the status line above the preview after a cutout. If it says CPU, the tool is running IS-Net on your processor, single-threaded, and 10 to 30 seconds per photo is normal there. The GPU path needs all of these:

• Chrome, Edge, Opera or another Chromium browser, not Safari or Firefox
• a computer, not a phone or tablet
• WebGPU with the shader-f16 feature (check chrome://gpu, which lists WebGPU status and features)
• at least 4 GB of memory, if the browser reports it

If you have all four and still see CPU, the GPU session probably failed to start and the tool fell back. Open the developer console: a line starting [background-remover] WebGPU model failed gives the reason. Updating the browser and the graphics driver fixes most of these. Models and hardware covers the rules in full.

It says GPU, but it is stuck or very slow. What now?

Older or integrated GPUs can run the model slowly, and a GPU that is also driving a busy desktop can stall. Things to try, cheapest first:

1. Close other GPU-heavy tabs (video, games, 3D) and try again.
2. Reload the page. A lost GPU device is recreated from scratch.
3. If one photo fails, select it and click Try again. A GPU run that fails or returns an empty mask switches the rest of the session to the CPU model on its own.
4. Try a smaller photo. Very large originals mean more work at full resolution after the model runs, especially with AI upscale on.

Why is there a dark or white halo in the hair?

Semi-transparent hair pixels still carry the old background's colour. The tool recovers the true colour of those pixels, which removes most halos, but a strong coloured backdrop or heavy motion blur can leave a trace. Check on black and on white with Background → Color, and if you can, use the GPU path, where BiRefNet gives cleaner hair than IS-Net. Edge quality explains what the colour recovery does and does not do.

Why were cream or white specks left behind?

The model was unsure about patches of background that resemble the subject, such as a cream wall behind a white shirt. Detached specks smaller than 2% of the subject are removed automatically, and faint haze away from the subject is cleared. Specks that survive are either attached to the subject or larger than that. The CPU model is more prone to this than the GPU model. Cropping the photo closer to the subject before you add it also helps, because the model then spends its resolution on the subject.

Why is part of my subject see-through?

Because the model thinks it is. Glass, smoke, veils, lace and thin fabric are partly transparent in real life and come back that way. A solid subject that turns partly transparent usually matches its background too closely in colour and texture. There is no manual brush in the tool; for a stubborn case, a photo with more contrast between subject and background gives the model more to work with.

Does it work on a phone?

Yes, with the CPU model, so expect 10 to 30 seconds per photo, sometimes more on an older phone. A few phone-specific points:

• iPhone HEIC photos open directly.
• Photos above 40 megapixels are refused with "This image is 48 MP. Resize it below 40 MP first" (with your photo's number). Resize with the image resizer first.
• Big batches use a lot of memory; if the tab reloads itself, the phone ran out. Work in batches of ten or so.
• AI upscale is not available on phones.

Why do Copy or Download fail?

Copy writes a PNG to the clipboard through the browser's clipboard API. Some browsers do not allow image writes, and some ask for permission. If the write is refused you see "Copying images is blocked in this browser. Download the image instead." In browsers with no image clipboard support at all, the Copy button does nothing. Download works everywhere.

If a download fails you see "The image could not be saved. Try a smaller size or another format." That usually means the browser could not create a canvas that large or encode it. The tool already caps output at 16.7 megapixels, so reduce the size under Resize or switch from PNG to WebP or JPG.

A failed ZIP shows "The ZIP could not be created. Download the images one at a time instead." It happens when memory runs out while packing a large batch.

What if the model fails to download?

You will see "Could not download the AI model (…). Check your connection and try again.", with the underlying reason in brackets. The model is served from huggingface.co, and downloads there redirect to Hugging Face's file storage hosts. A network filter, VPN, corporate proxy or ad blocker that blocks those will stop it. Allow huggingface.co and the host it redirects to, or try another network. Nothing about your photo is involved; it is a plain file download.

What do the other error messages mean?

Message	Meaning	What to do
"This file could not be read as an image."	The browser could not decode it	Check the file opens elsewhere, or convert it with the format converter
"This is a PDF. Convert the page to an image first, then remove its background."	The file is a PDF, whatever its name says	Use PDF to PNG first
"The model returned an empty or invalid mask."	A run produced no usable mask	On the GPU path the tool retries on the CPU automatically; otherwise click Try again
"This browser cannot draw images off the main thread."	No OffscreenCanvas support	Update the browser
"The background remover could not start in this browser."	The worker failed to start	Update the browser, or try Chrome, Edge, Firefox or Safari
"Daily limit reached for this tool."	The guest usage limit for today is used up	Sign in or see pricing, or come back tomorrow
"The sample image could not be loaded. Try one of your own."	The sample photo host was unreachable	Use your own photo

Still stuck? Contact us with the browser, device, the status line text and any console message, and we will look into it.