What changed

upscale-image previously called a fal.ai ESRGAN model that returned an upstream 422 on well-formed requests, so paid calls did not produce an image. It now points at the Venice image/upscale backend (venice-sd35) — the same handler its sibling image-upscale already uses in production — and returns a permanent hosted URL.

What this means for callers

  • Send image_url (a public http/https image) and an optional scale of 2 or 4 (default 2).
  • Earlier requests advertised a 1–8 scale range that the backend did not honor; the input schema now reflects the real 2×/4× support, so a request that validates is a request the backend can serve.
  • Output includes the upscaled image_url, the scale, and the model.