AI Image Denoiser

Every camera sensor produces some random electrical signal alongside the real image. In bright daylight that signal is drowned out and photos look clean. In dim light the camera amplifies everything to compensate, and the random part gets amplified too. That is the colorful speckle you see in night shots, indoor photos, and anything taken at high ISO settings.

Small sensors suffer most, which is why phone photos show noise sooner than photos from large cameras. Older digital cameras, cheap webcams, and heavily brightened shots all share the same problem. Compression adds its own layer on top: JPEG saves space by discarding detail, and at aggressive settings it leaves faint blocky patterns that behave much like noise.

None of this means the photo is lost. The underlying image is still there beneath the speckle, and separating the two is precisely the job a trained denoising network does well.

The classic way to hide noise is to blur the image, averaging each pixel with its neighbors. Noise fades, but so does everything else: edges soften, texture disappears, and the photo takes on a plastic look. Blur cannot tell noise from detail because it does not know what either looks like.

This tool runs SCUNet, a network trained on pairs of clean and realistically degraded photos. It learned the statistical difference between random speckle and real structure, so it can remove one while preserving the other. Fabric keeps its weave, skin keeps its pores, and edges stay where they are, while the grain between them disappears.

It is also blind denoising, meaning you do not need to measure or guess a noise level. The network estimates the degradation in each region of the photo on its own, treating a clean sky and a noisy shadow differently within the same image. Large photos are processed in overlapping tiles, so resolution is preserved from corner to corner.

The denoising model downloads once, about 73 MB, and is stored by your browser. From then on every photo is processed locally on your own hardware. Nothing is uploaded, no queue, no account, and no per-image cost, which is exactly why there is no watermark or daily limit either.

With a modern graphics chip the browser runs the network through WebGPU and a typical photo finishes in a few seconds. Without one, the processor fallback is slower but produces the same output. Photos up to 4,096 pixels on the longest side are accepted, and the output keeps the original dimensions exactly, only cleaner.

The output format is your choice. PNG preserves the cleaned result losslessly, while JPEG and WebP produce smaller files. If storage is the bigger concern, cleaning the photo here and then shrinking it with the Image Compressor is an effective pair, noise compresses poorly, so denoised photos also produce smaller files at the same quality.

Denoising shines on grain, speckle, and mild compression artifacts. A dim restaurant photo, a grainy scan, or an old phone picture typically comes out looking like it was shot with a much better camera. Colors settle, surfaces calm down, and the subject stands out again.

The limits are worth knowing before you judge a result. Detail that noise fully destroyed cannot be recovered, so areas that were pure speckle become smooth rather than detailed. Extremely strong noise costs some fine texture even in a good result, that trade is inherent to the problem. Motion blur and missed focus are different defects entirely and do not improve here.

Order matters when combining tools. Always denoise before enlarging with the AI Image Upscaler, because upscaling first turns every speck into a larger artifact. For old portraits where the face itself is the problem, follow the denoiser with the AI Face Restorer, which rebuilds facial detail the noise took away.