Classic resampling is not AI upscaling

Classic resampling — Lanczos, bicubic, bilinear — interpolates existing pixels and creates no new detail. Upscale a 200 × 200 pixel image to 1000 × 1000 and you get a mushy image. More on that in our scaling guide.

AI upscaling does something different: a neural network was trained on millions of "small/large" pairs and learned what likely detail would be at the higher resolution. The result is hallucinated detail — not present in the original pixel grid, but statistically plausible.

The models in 2026

  • Real-ESRGAN (2021, further developed 2024): the open-source standard. Runs browser-local via ONNX, speed acceptable with WebGPU. Best general quality for photo content.
  • ESRGAN-Anime / waifu2x: specialized in anime, manga, pixel art. Excellent for line content, fails on real photos.
  • Topaz Gigapixel AI: commercial, €99 one-time. Best available quality for photo restoration. Local, GPU-accelerated.
  • BSRGAN (Microsoft Research, 2021): an open-source alternative to Real-ESRGAN with a different training pipeline. Often better on text, comparable on photos.
  • SUPIR (2024): a generative variant that works with diffusion models. Impressive, but prone to hallucinations — a face can become a different person.

When AI upscaling is the right choice

  • Photo restoration of old shots. Family photos from the '80s originally scanned at low resolution. Real-ESRGAN or Gigapixel often deliver impressive results.
  • Web-image upgrades for retina. Existing web assets originally meant for 96-dpi displays, brought to retina quality — when the original is no longer available.
  • Anime/manga enlargement. waifu2x beats any classic resampling.
  • Stock-photo upgrades. A purchased stock photo at 800 × 600 is too small for a print ad; Gigapixel rescues it.
300 × 250 — Rectangle
Cookie-Banner ausstehend

When AI upscaling is the wrong choice

  • Forensic identification. AI upscaling hallucinates — the resulting face isn't the original's. Law enforcement mustn't use the results.
  • Documents and text in untrained fonts. Real-ESRGAN was trained primarily on photo data. With rare scripts (italic, Asian, handwritten) pseudo-characters emerge.
  • Images with fine patterns (fabric, wood, marble). Where classic bilinear simply softens, Real-ESRGAN visibly hallucinates wrong patterns.
  • At extreme enlargement (10×+). Most models are trained for 2×–4×. At 10× the result becomes increasingly "surreal".

Browser-local upscaling 2026

Real-ESRGAN runs reliably browser-local in 2026 via onnxruntime-web and WebGPU. Performance:

  • 500 × 500 → 2000 × 2000: ~3 seconds on an M2 with WebGPU.
  • 1000 × 1000 → 4000 × 4000: ~15 seconds.
  • 2000 × 2000 → 8000 × 8000: often >1 minute, memory becomes critical.

Browser-local upscaling has two strategic advantages: privacy (see our background-removal post) and no subscription lock-in. Model sizes of 50–200 MB are downloaded on the first visit and cached.

Hybrid workflow: AI + classic sharpening

Pro tip: AI upscaling sometimes produces soft results. After upscaling, a classic unsharp mask pass (Photoshop, Lightroom) often delivers the best end result. Real-ESRGAN itself has "GAN" variants (with additional sharpening), but that also amplifies artifacts. Better practice: conservative Real-ESRGAN + manual sharpening with control over strength and radius.

300 × 250 — Rectangle
Cookie-Banner ausstehend

Comparison to Lanczos

For web delivery in the 2×–3× range on high-quality photo content, Real-ESRGAN delivers measurably better results than Lanczos. For UI sprites and pixel art, however, Lanczos is often the right choice — AI models interpret pixel lines as edges to be smoothed.

Decision rule:

  • Photo content, 2×–4×: Real-ESRGAN.
  • UI icon, pixel art, 2×–4×: Lanczos.
  • Logo (vector available): never upscale — use SVG.
  • Anime, manga: waifu2x.
  • At extreme enlargement (4×+): Topaz Gigapixel or SUPIR with caution.

File-size implications

Upscale an image and you get a bigger file. A 4000 × 4000 AVIF is typically 6–10× larger than a 1000 × 1000 AVIF at the same quality. That makes upscaling on the web sensible only when the higher resolution visibly adds value — such as retina display or print preparation. More on format optimization in our AVIF vs. WebP vs. JPEG XL post.

Honest limitations in 2026

AI upscaling is astonishingly good in 2026 — but no miracle cure. A too-heavily-compressed JPG (quality 30) can't be restored to "original quality" even with Real-ESRGAN. The block artifacts are partly concealed, but the image will never reach the result achievable from the original RAW.

To have a valuable photo in the best possible quality: keep the original RAW, export from it to a higher resolution as needed. AI upscaling is a rescue, not a substitute for proper archiving.

300 × 250 — Rectangle
Cookie-Banner ausstehend

Tooling recommendation 2026

  • One-time, all photo: Topaz Gigapixel (€99, local).
  • Open source, photo: Real-ESRGAN via chaiNNer (GUI) or the command line.
  • Browser-local, no installation: Real-ESRGAN web apps with onnxruntime-web.
  • Anime and pixel art: waifu2x.
  • Mass pipeline: Real-ESRGAN Docker in CI/CD.

Sources

Real-ESRGAN on GitHub · Topaz Gigapixel AI · BSRGAN · waifu2x · SUPIR · chaiNNer (GUI) · ONNX Runtime Web · Real-ESRGAN Paper (arXiv).