What does AI Image Enhancer do?
Free AI image upscaler and enhancer: enlarge photos 2x, 4x or to 4K and 8K with Real-ESRGAN, fixing blur and JPEG artefacts, in your browser. No uploads.
- Price
- Free
- Account
- Not required
- Processing
- Entirely in your browser; your data and files are not uploaded
- Works on
- Any modern browser on desktop, tablet or phone
- Category
- Image Tools
How to use the AI Image Enhancer (Upscaler)
- Drop a JPG, PNG or WebP image of any file size onto the drop zone.
- Choose the output size: 2x, 4x, Full HD, 4K or 8K. The tool shows the exact output dimensions and warns if a size is too large for your device.
- Choose the AI detail: Maximum detail (faithful, restores real fine texture, needs WebGPU), Standard (fast), or Creative detail (the AI redraws fine details such as petal texture; the image is sent to Cloudflare Workers AI).
- Pick the enhancement level: Natural (recommended), Strong, or Off for the pure AI result. Choose "Same size" to only enhance without enlarging.
- Select Enhance. The first time, the AI engine (about 28 MB) and the model (34 MB for maximum detail, 4.6 MB for standard) download and are cached by your browser.
- Compare with the slider. Switch to 100% (actual pixels) to inspect detail, then download a PNG (lossless) or a much smaller JPG.
Worked example
A soft 610 x 396 photo to 4x and 8K
At 4x the output is 2,440 x 1,584 pixels, processed as 12 overlapping tiles. At 8K the long side becomes 7,680 pixels (7,680 x 4,986) using two AI passes. In our test on a desktop GPU the 4x result took about 4 seconds including the model load and the 8K result about 32 seconds. On a test image with known detail, edge sharpness of the 4x result was about 3.4 times that of a normal browser enlargement and close to the real high-resolution image.
How it works
Maximum detail runs Real-ESRGAN x4plus and Standard runs Real-ESRGAN general x4 v3, both super-resolution networks trained on realistically degraded photos, so they also reduce blur, noise and compression artefacts. They run with ONNX Runtime Web in your browser (WebGPU, or WebAssembly on the CPU for Standard). The model files are served from this site; the browser checks each file's SHA-256 before use, and both were checked against the publisher's reference outputs. Creative detail first sends a copy of the image, resized to about one megapixel, to FLUX.2 [klein] 4B by Black Forest Labs on Cloudflare Workers AI, with an instruction to keep the composition, subject and colours and add fine detail at the chosen creativity; the result is then upscaled in the browser. The image is split into overlapping tiles; each tile is enlarged 4x and cross-faded into the result along the overlaps so no seams show. Sizes other than 4x are reached with high-quality resampling of the 4x result; for more than 4x (for example 8K from a small photo) the first result is resampled to a quarter of the target and enlarged 4x again. Transparency is enlarged with smooth resampling. An optional finishing step applies auto levels (the darkest and brightest 0.5% of pixels are mapped to black and white with one curve for all channels), a gentle vibrance boost and light unsharp-mask sharpening.
Assumptions
- The model predicts detail from patterns it learned during training. It does not know what was really in the scene.
Frequently asked questions
Are my images uploaded?
No. The model runs on your device. Only the AI engine (from jsDelivr) and the model (from this site) are downloaded; your image never leaves your browser.
Can it make a 4K or 8K image?
Yes, with WebGPU (current Chrome, Edge and many other browsers). 4K means 3,840 pixels on the long side and 8K means 7,680. Very small images need two AI passes for 8K, and the tool allows at most 16x enlargement and about 40 megapixels of output so it fits in browser memory.
Can it restore a blurry photo or make a face recognisable?
It removes mild blur, noise and JPEG artefacts and makes edges crisp, but it cannot recover information that is not in the original. Do not rely on it for identification or evidence.
Which AI detail mode should I use?
Maximum detail is the most faithful: it restores real texture that survives in the photo and does not invent new objects. Creative detail is for images that are too soft for that, such as small illustrations: it redraws fine details like petal texture, stamens or leaf veins, so the result looks crisper but differs from the original. Do not use Creative detail for documents, evidence or faces that must stay accurate.
Is my image uploaded in Creative detail mode?
Yes, only in that mode: a copy resized to about one megapixel is sent to Cloudflare Workers AI to be redrawn and is not stored by us. The other modes run entirely in your browser.
Why does the result look similar at Fit size?
At Fit size both images are scaled down to your screen, which hides most of the extra detail. Switch the comparison to 100% (actual pixels) to see the difference.
Which models and licences are used?
Real-ESRGAN x4plus and Real-ESRGAN general x4 v3 by Xintao Wang (BSD 3-Clause licence, published with the models under /models/), and for Creative detail FLUX.2 [klein] 4B by Black Forest Labs, used through Cloudflare Workers AI under the Black Forest Labs terms of service.
Limitations
- Without WebGPU only the Standard model is available: it runs on the CPU, much more slowly, with one AI pass and about 4 megapixels of output.
- Creative detail invents details, can change small shapes and colours, has a daily limit, and needs the site to run on Cloudflare.
- Maximum 16x enlargement and about 40 megapixels of output (8K at a 3:2 shape). Large outputs need a few hundred megabytes of free memory.
- An 8K PNG can be over 30 MB; use the JPG download for a much smaller file.
- It can invent texture and artefacts, especially on text, faces and fine patterns. Always compare with the original.