Agrandissez avec IA

Classical upscaling algorithms — bicubic, bilinear, Lanczos — work by interpolating values between existing pixels. They have no knowledge of what the image "should" look like at higher resolution, so they produce smooth but blurry results. Edge sharpness is lost, fine textures become mushy, and text or line art develops halos. The larger the scale factor, the worse the result.

Real-ESRGAN is a Generative Adversarial Network trained on millions of real-world high-resolution images alongside artificially degraded (downsampled, blurred, compressed) versions. The generator learns to map low-resolution input to high-resolution output that is statistically indistinguishable from genuine high-res photos. It does not merely interpolate — it hallucinates plausible detail based on learned patterns in textures, edges, and structures.

The standard Real-ESRGAN model is optimised for photographs with natural textures (skin, grass, fabric, architecture). It adds grain and texture detail that matches photo characteristics. The Anime6B model is optimised for drawn content: flat colour regions stay flat, linework stays crisp, and the model avoids adding photographic noise to illustrated elements. Always choose the model that matches your content type for best results.

AI upscaling is computationally expensive. Processing a 2000×2000 image at 4x produces an 8000×8000 pixel (64 megapixel) output — equivalent to a medium-format camera. We cap free inputs at 2000×2000 to keep processing times reasonable. If your input is larger, resize it to 2000px on the longest edge first using our Image Resizer, then upscale.

For photographs: remove compression artefacts first by running through the Image Compressor at quality 95 to clean up JPEG blocking before upscaling. For game sprites: use 4x with the anime model and ensure your input is at the original pixel-art resolution (no prior bicubic scaling). For print output: upscale to at least 300 dpi for your intended print size — a 4-inch print needs 1200px minimum.