What is Image Upscaling?

Bicubic interpolation is a fixed math formula: it looks at the 16 nearest pixels, fits a smooth curve through their values, and samples new pixels along that curve. It has no idea what it is looking at, so it treats the edge of a logo the same way it treats a soft gradient in a shadow, which is why bicubic-resized images look soft and slightly waxy once you push past about 150% of the original size. AI upscaling models (Real-ESRGAN and its predecessors were the architectures that popularized this, and most commercial tools including Dezygn's now run proprietary variants) instead learn a mapping from low-resolution patches to high-resolution ones during training, using millions of image pairs where the model is shown a degraded version and scored on how close its reconstruction gets to the real high-resolution original. Applied to a product photo, the model does not stretch pixels; it re-synthesizes edges, stitching, grain, and text based on patterns it has seen before. That is also the source of its main risk: on unfamiliar textures or heavily compressed source files, the model can invent detail that was never there, which shows up as slightly artificial-looking sharpening or repeating micro-patterns in fabric.

In a product-photography workflow, when you upscale matters as much as whether you do. Upscaling after generation, meaning taking a finished AI-generated or photographed image and enlarging it for a specific placement (a hero banner, a print asset, a marketplace zoom requirement), is the safer and more common use case, because the composition and lighting are already locked and the model only has to add resolution, not invent content. Upscaling before generation, meaning cleaning up a low-quality source photo before feeding it into an image-to-image model as the reference for a new scene, is riskier and matters more than most sellers assume. Image-to-image generation is conditioned on the source: a blurry, low-resolution, or heavily compressed product photo passes its softness and artifacts into every generated variation, and no amount of downstream polish fully removes that inherited fuzziness. A rule of thumb: upscale the source first if it looks visibly soft at 100% zoom, then generate; if the source is already sharp, upscale the finished output only if the destination (print, a large hero placement, a zoom-enabled listing) actually requires more pixels than you started with.

On magnitude, 2x upscaling (doubling both dimensions, so 4x the pixel count) is close to lossless for almost any reasonably sharp source and is the default choice when the goal is simply hitting a platform's minimum resolution, such as Amazon's 2,000-pixels-longest-side requirement for zoom. 4x upscaling asks the model to invent far more detail per pixel and is where artifacts become visible: watch for over-sharpened, slightly crunchy edges around text and logos, waxy or plasticky skin and fabric texture, repeating patterns in areas the model treats as background noise (mesh, knit, gravel), and color banding in smooth gradients like studio backdrops. The practical fix when 4x looks artificial is not to abandon upscaling but to step it: upscale 2x, inspect, and only push to 4x if the source genuinely supports it. As a general guide, an image that is 20 to 50 percent below target resolution upscales cleanly; an image that needs to be quadrupled in linear dimension will show some artifacts under close inspection even with strong modern models, though most product categories still tolerate this better than portraits or faces do.