What a JPEG artifact remover actually removes
This JPEG artifact remover runs the same models as the free AI image upscaler, but it starts from the job instead of the size: cleaning an image that has been compressed too hard. JPEG saves space by splitting a picture into 8 x 8 pixel squares and throwing away fine detail in each one. Push the quality setting down, or save the same file several times, and three kinds of damage appear: visible blocks in skies and smooth areas, ringing or mosquito noise as a fuzzy halo around text and sharp edges, and banding where a gradient breaks into steps. The default model here, GAN, treats the blocks and halos as damage and paints them out while it enlarges the image 4x, so smooth areas come back smooth and edges come back clean. Real-ESRGAN was trained on exactly that kind of degradation too and gives a slightly softer result.
It is different from deblurring. If a photo is soft because of camera shake or missed focus rather than compression, the free unblur image tool is the better fit.
How to remove compression artifacts from a JPG
Upload the JPG, or a WebP or PNG that was saved from a JPEG, up to 2048 px per side. Press Remove Artifacts and wait a few seconds. The result appears next to the original on a slider: zoom into a sky, a wall, or the edge of some text, where artifacts are easiest to see, and compare.
Pick the model by the result you want. GAN is the default and gives the cleanest, crispest result on most compressed photos, screenshots, and graphics. Real-ESRGAN is a little softer, which can suit faces and skies. Seed VR 2 Small is a diffusion restorer built for blur rather than compression: on a heavily blocked file it tends to read the blocks as texture and keep them, so it is the last choice here.
To see whether a file has artifacts at all, zoom to 200 or 300 percent on a smooth area such as sky, skin, or a painted wall. A grid of faint squares, colors that change in steps, or a shimmering halo hugging the edge of text are the signs. Images that look fine at that zoom have nothing for this tool to remove, and the general upscaler is the better choice if you only want them bigger. Screenshots of photos, images saved from messaging apps, and pictures cropped out of web pages are the usual suspects, because each of those steps recompresses the file.
Fix JPEG quality, then pick the final size
GAN and Real-ESRGAN clean and enlarge in the same pass. GAN returns 4x the width and height of the input, and Real-ESRGAN does the same for small images and returns larger ones at up to about 16 megapixels. That is useful, because compressed images are often too small as well. If you only wanted the artifacts gone and the original size back, shrink the cleaned result with the free image resizer. Downscaling a clean 4x image also gives a slightly sharper result than the original ever had at that size.
When you save the fixed image, choose PNG or a high JPEG quality setting, around 90 or above. Saving it again at a low quality brings the blocks straight back. The download from GAN or Seed VR 2 Small is a high-quality WebP, which keeps the cleaned detail while staying a reasonable file size, and Real-ESRGAN returns a lossless PNG.
When artifacts are too far gone
Compression throws information away for good. The models can hide the damage convincingly and restore edges, but they cannot recover the exact detail that was discarded. A face that is a handful of blocks wide, or text that compression has merged into smudges, will come back cleaner but not as it originally was. Start from the oldest, largest copy you can find, such as the original upload rather than a forwarded or screenshotted version, because every re-save adds new artifacts on top of the old ones.
Typical wins are old web graphics, product photos from marketplace listings, images downloaded from messaging apps, scanned documents saved as low-quality JPEGs, and memes that have been shared until they look deep fried. Your images are processed transiently and not stored.
If the image also has a watermark or a date stamp you are allowed to remove, clean the artifacts first and remove the overlay second, since the removal tools rebuild the background better from a clean source.