Interpolation is the process of generating new pixel values from existing ones, used whenever an image is resampled, enlarged, rotated, or demosaiced from sensor data. Every raw file passes through interpolation during demosaicing, since the sensor’s Bayer filter records only one color per photosite; the other two colors at each location are inferred from neighboring values. Enlargement, downsizing, and any geometric transformation also rely on interpolation to fill the new pixel grid.
Several algorithms exist, ordered roughly by computational cost. Nearest-neighbor picks the closest existing pixel, producing blocky results but preserving hard edges, useful for pixel art. Bilinear averages the four nearest pixels, producing smoother but blurrier results. Bicubic samples 16 pixels and applies a weighted cubic function, balancing sharpness and smoothness; it has been the default in Adobe Photoshop for decades. Lanczos uses a sinc function and a larger sample window, producing crisper results for photographic content and is preferred in video and high-end print workflows.
AI-based interpolation has reframed the field. Lightroom’s Super Resolution feature, ON1 Resize AI, Topaz Gigapixel, and the upscale modes in Photoshop all train neural networks on large datasets of paired low and high-resolution images and learn to hallucinate detail that classical interpolation simply averages away. The results can look extraordinary on faces, fabric, and foliage, but the model is filling in what it expects rather than recovering what was photographed. On unfamiliar subjects or fine repetitive patterns, AI upscalers can invent artifacts that look plausible but are not in the original scene.
Demosaicing is the most consequential interpolation step most photographers never see. The Bayer pattern on a typical sensor places one red, one blue, and two green filters in every 2×2 block of photosites, so 75% of each color channel is interpolated rather than measured. Different demosaicing algorithms (AHD, AMaZE, DCB, LMMSE, and others) trade off color accuracy, edge sharpness, and moire suppression. Capture One, Lightroom, RawTherapee, and DxO each ship their own engines, which is why the same raw file looks subtly different in different programs.
Sampling theorems set the limits. Doubling resolution doubles the unknown pixel count at each step, and beyond about 200% to 400%, classical interpolation produces noticeable softness and aliasing. AI methods extend the useful range but cannot recreate information that was never recorded; a tiny patch of distant text remains illegible even after a 4x upscale, just with cleaner-looking jaggies. Pairing high-quality capture with conservative resampling generally outperforms pushing low-quality capture through aggressive AI enlargement.
Practical advice for prints and exports: downsample with bicubic sharper or Lanczos for the final size, upsample modestly with bicubic smoother or AI tools for large prints, and avoid repeated interpolation across edits. Each pass loses a small amount of fidelity, so cropping, rotating, and resizing in a single non-destructive step keeps the file cleaner than chaining them through multiple saves. The sharpening applied after interpolation is also part of the chain and deserves to be tuned for the final output size.