Upscale images 2x-4x with Real-ESRGAN and remove backgrounds with U2-Net in one Actor. Bulk URLs in, downloadable PNG/WebP/JPEG out, one dataset row per image. Pay only for successfully processed images.
First version, built and tested locally. Not yet published.
Upscale 2x/3x/4x with Real-ESRGAN realesr-general-x4v3 on CPU (onnxruntime), with adjustable denoise strength (blend with realesr-general-wdn-x4v3).
Background removal with U²-Net, transparent or solid-color background.
Combined "remove background, then upscale" operation, and a square-canvas layout for marketplace-style product images.
Inputs from image URLs (strings or {url} objects) and Apify key-value store keys.
Outputs in PNG, WebP and JPEG to the run's key-value store, plus one dataset row per image with dimensions, timings and stable error codes, and an OUTPUT run summary.
Pay-per-event charging after the result is saved (image-upscaled per started 0.25 MP of input, background-removed per image). Respects the user's max cost per run, with in-flight reservation so parallel images cannot overspend.
Robustness: download retries with backoff (honors Retry-After), timeouts, content-type and size checks, private-network URL blocking, header-only size check before decoding, memory-aware tiling, per-image errors that never fail the run, and migration-safe progress state.
Model weights are baked into the Docker image at build time from the official release URLs, SHA-256 verified.
0.1.3 (2026-10-23)
A run where nothing succeeds now ends as FAILED (it used to show green with only error rows), so schedules and monitoring see real outages. Nothing is charged for failures, as before.