Real-ESRGAN: What a Super-Resolution Model Can Enhance—and What It Cannot Recover

Real-ESRGAN is a practical super-resolution and restoration project introduced in 2021. It can produce sharper, higher-resolution output, but generated texture should not be mistaken for details physically recorded in the original image.

The research and software

The original repository cites the 2021 paper Real-ESRGAN: Training Real-World Blind Super-Resolution with Pure Synthetic Data. It provides model variants, inference scripts, tiled processing and integrations with face enhancement systems. Some implementations can process general photographs, while others are optimized for anime illustrations. Every model choice should be tracked explicitly rather than labeled generically “4x upscale.”

Start by classifying the input

A clean but small digital illustration is different from a heavily compressed JPEG or a faded archival photograph. Inspect blur, ringing, noise and existing pixel structure at native size. Keep a lossless original, then choose an appropriately documented model and conservative scale. Upscaling cannot reconstruct details that were never recorded; it predicts textures consistent with learned examples.

A practical restoration review

For each test image save an original and output side by side at equivalent display dimensions. Inspect lettering, railings, fabric weave, jewelry, thin branches and faces; these can all appear convincingly wrong. Confirm that any crop or alpha transparency survives. Tiled inference can change seams, and fractional output scaling might include ordinary resizing after the neural stage.

Avoid false documentary confidence

Historical photos are sensitive because generated signs, uniforms or architectural details may be interpreted as facts. Keep the source scan and processing recipe, and use a caption noting AI-assisted restoration when it materially changes detail. Do not overwrite catalog masters or use a reconstruction to authenticate text or identify someone. A forensic pipeline needs a stricter evidence-preservation standard.

Operational and license caveats

The upstream project describes code and available weights; derivative builds, model packs and hosting services may impose different legal and safety rules. Even when source code is permissively licensed, review any weights and input rights separately. Never download binaries from anonymous mirrors when the original maintained repository can be verified. Local GPU memory and performance depend on tile size and image dimensions.

Alternatives and scope

Use traditional resampling for clean vector-like content where invented detail would be inappropriate. For face-specific degradation, a specialized face-restoration model may offer a better perceptual result but carries stronger identity risk. For editorial proof or archiving, faithful scanned originals remain authoritative, and an AI-upscaled version should be treated as a derivative presentation asset.

Worked scenario: a hypothetical production brief

Example: A museum receives a 640-pixel scan of an old printed railway timetable. A neural upscaler produces attractive sharper letters, but the output could have invented strokes inside tiny numbers. The curator retains the scan, links to the historical source and uses the upscaled image only for a clearly labeled visitor display. Any searchable dates are transcribed from the original or an independent copy, never from reconstructed pixels alone. For precise public information, faithful transcription is more valuable than a falsely legible image.

Additional editorial and quality safeguard

When the target is print, inspect the enlarged image at expected print dimensions and viewing distance rather than only at enormous zoom. A mild interpolation that appears less spectacular at 400 percent may reproduce the original information more faithfully. For maps, charts and labeled diagrams, recover exact marks from the source document instead of trusting neural texture synthesis.

Sources & Further Reading

• Tencent ARC / Xintao Wang — Real-ESRGAN official repository (living documentation; version pin required)
• Tencent ARC — GFPGAN official repository (living documentation; version pin required)