🤖 Model 🖼️
GFPGAN — AI Face Restoration & Enhancement Model
GFPGAN by TencentARC is a practical face restoration model that enhances and restores degraded face images. It leverages...
← All guides
Face Restoration API
A face restoration API takes a blurry, damaged, low-resolution, or AI-generated face and returns a clean, sharp, natural-looking one. The two search intents behind it are old-photo repair (family photos, scans, archives) and fixing the slightly-off faces that image generators still produce.
On Replicate this space has unusually clear leaders: GFPGAN (115M+ runs) and CodeFormer (54M+ runs) are two of the most-used models on the entire platform, with Real-ESRGAN (94M+ runs) as the general upscaler that can optionally enhance faces along the way.
Quick Recommendation
Start with tencentarc/gfpgan. It is fast (typically a couple of seconds per image), battle-tested, and the default choice for both old photos and AI-generated faces.
Switch to sczhou/codeformer when GFPGAN's output looks too smooth or too different from the person — CodeFormer's codeformer_fidelity dial (0–1) lets you trade restoration strength against staying faithful to the original face.
Use nightmareai/real-esrgan when the whole image needs upscaling and faces are only part of the job — its face_enhance flag runs face correction on top of general upscaling in one call.
Keyword-Matched Use Cases
| Search intent | What the user likely needs | Best starting point |
|---|---|---|
| Face restoration API | One call: damaged face in, clean face out | GFPGAN |
| Restore old photos | Faces plus scratches, grain, fading | GFPGAN, then colorize or upscale |
| Fix AI-generated faces | Repair artifacts from SD/Flux outputs | GFPGAN or CodeFormer at high fidelity |
| Unblur face photo | Deblur and sharpen a soft portrait | CodeFormer (tune fidelity) |
| Photo enhancer API | Whole-image upscale incl. faces | Real-ESRGAN with face_enhance |
| Old photo restoration pipeline | Restore, then colorize, then upscale | GFPGAN → colorizer → light upscale |
Model Comparison
| Model | Runs on Replicate | Typical speed | Key control | Best for |
|---|---|---|---|---|
| tencentarc/gfpgan | 115M+ | ~2s | version (v1.3 / v1.4), scale |
Default face restoration, old photos, AI faces |
| sczhou/codeformer | 54M+ | Slower per run | codeformer_fidelity 0–1, background_enhance, face_upsample |
Controllable restoration when identity must be preserved |
| nightmareai/real-esrgan | 94M+ | ~3s | scale, face_enhance |
Whole-image upscaling with optional face correction |
How to choose between GFPGAN and CodeFormer
GFPGAN is the speed-and-simplicity pick. Three inputs (img, scale, version), fast responses, and results that look good on the vast majority of portraits. Its weakness: on heavily degraded faces it can hallucinate — the restored face looks clean but subtly unlike the person.
CodeFormer exists precisely for that failure mode. Its codeformer_fidelity parameter is a dial between quality and identity: low values restore aggressively (cleaner but more invented detail), high values stay close to the input (more faithful but less repaired). It also bundles background_enhance and face_upsample, making it closer to a one-call photo enhancer. The tradeoff is speed — expect noticeably longer runs than GFPGAN.
Real-ESRGAN is not primarily a face model. It is the standard general-purpose upscaler on Replicate, and its face_enhance option applies GFPGAN-style face correction during upscaling. If your input is a whole photo that needs 2–4× resolution and the faces just need to survive the process, one Real-ESRGAN call beats a two-model pipeline.
Restoration pipelines
Face restoration is often one step in a larger old-photo workflow:
- Denoise heavily degraded scans first — SwinIR handles grain, compression artifacts, and scan noise.
- Restore faces — GFPGAN or CodeFormer.
- Colorize black-and-white photos — see our AI photo colorizer API comparison.
- Upscale the final result — Real-ESRGAN, or skip if face_enhance already ran.
Running face restoration before colorization gives the colorizer cleaner input and avoids amplifying artifacts into false color patches.
🤖 Model 🖼️
sczhou/codeformer
Restore and enhance faces in old photographs and AI-generated images. Takes a degraded or low-quality face image as inpu...
🤖 Model 🖼️
nightmareai/real-esrgan
Upscales images using Real-ESRGAN with adjustable scaling factors from 0 to 10x. Includes optional GFPGAN face enhanceme...
🤖 Model 🖼️
jingyunliang/swinir
Restores and enhances images using Swin Transformer architecture. Performs real-world image super-resolution in large an...
Practical Decision
Default to GFPGAN — it is fast, cheap per run, and right for most inputs. Offer CodeFormer as the "stronger restoration" option, exposing the fidelity slider if your users care about likeness. Use Real-ESRGAN with face_enhance when the job is really an upscale that happens to contain faces.
All three are pay-per-run on Replicate with no idle cost, so for low volume there is no reason to self-host.
FAQ
What is the best face restoration API?
GFPGAN is the most widely used face restoration model on Replicate (115M+ runs) and the best default. CodeFormer is the strongest alternative when you need to control how faithful the restored face stays to the original.
GFPGAN vs CodeFormer — what is the actual difference?
GFPGAN is faster and simpler; CodeFormer is slower but adds a fidelity dial (0–1) plus background enhancement and face upsampling. If GFPGAN output looks too smooth or changes the person's appearance, switch to CodeFormer with fidelity around 0.7.
Can these fix AI-generated faces?
Yes — both GFPGAN and CodeFormer were designed for old photos and AI-generated faces. They are commonly chained after Stable Diffusion or Flux generations to clean up eyes, teeth, and skin artifacts.
Should I restore faces before or after upscaling?
Restore first, upscale second — or do both in one call with Real-ESRGAN's face_enhance. Upscaling a damaged face first only enlarges the artifacts the restorer then has to fight.