GFPGAN
Free.ai (self-hosted)
·
image_enhance
·
~1000 tokens per image
GFPGAN is an image upscaler built by Tencent. Strongest at Portrait restoration, photo touch-up.. Self-hosted on Free.ai GPUs — runs free against your daily token pool (1,000 tokens per image). Released under Apache 2.0 — commercial use permitted on Free.ai.
Similar models
Frequently Asked Questions
GFPGAN upscales and sharpens low-resolution images while preserving detail. Upload a small or blurry photo and GFPGAN returns a higher-resolution version suitable for print, large displays, or further editing.
2× and 4× are the standard targets; some configurations support 8×. Final size is capped at 8192×8192 pixels on Free.ai to keep memory bounded.
PNG, JPG, WebP, HEIC, and BMP up to 10 MB. Animated GIFs are processed frame-by-frame.
Yes - modern enhancers include face-aware restoration. For very degraded portraits the dedicated /image/restore-face/ tool gives the cleanest result. GFPGAN handles general scenes plus faces in one pass.
GFPGAN runs on our own GPUs against your daily free pool first; then $5 → 200,000 paid tokens. About ~500 tokens per upscale.
GFPGAN hallucinates plausible detail at the upscale step - it is not a forensic tool. For most consumer uses (web → print, mobile → desktop wallpaper, archive scan → display) the result is convincing.
Crisp on large text + simple logos; small fine print can warp on aggressive (4×+) settings. Use 2× when text fidelity matters most.
Yes - /batch/ accepts a folder. Outputs land in /account/?tab=history with the original filename suffixed `-upscaled`. For folder-tree preservation use the API.
Yes - POST to /v1/image/enhance/ with model="GFPGAN" and your image. /api/ has the full reference + auth.
Same policy as the rest of Free.ai - self-hosted models stay on our GPUs; premium pass through with a DPA; uploads expire on the share-window schedule (24h anon, 7d signed-in). We do not train on your images.
Yes. Free.ai grants commercial use of upscaled images. You must own the source you uploaded.
5-20 seconds at 2×, 15-60 seconds at 4× depending on source resolution and GPU load. Premium models tend to finish on the lower end of those ranges.