Enhance a Face in a Photo — Pro, on our server
The face pass is a separate job from the clean-up, and it is deliberately narrow. After the whole photograph goes through a clean-up model, a detector looks for faces; each one it finds is cut out, restored by a model trained on faces specifically, checked against the version that arrived, and pasted back under a mask that follows the outline of the face rather than a box around it. Narrowness is the feature. A model given free rein over an entire portrait will happily redraw a hairline; a model confined to the skin inside a parsing mask cannot. The Cloud HD enhancer also never requires a face at all — a photograph with nobody in it simply gets the other passes.
Group photographs behave differently from portraits, and knowing that saves disappointment. Each face is treated on its own terms, so results within one frame vary honestly: the person nearest the camera gains a lot, the person at the back gains little, and a face turned too far away may be passed over entirely. That unevenness is the tool telling the truth about how much information each face carried. If a single person is what matters, crop tightly to them with the on-device crop tool before uploading, so the frame you send is mostly the face you care about.
Restoring a face is not the same as flattering one. Nothing here slims a jaw, whitens teeth, removes a blemish or applies the smoothing that phone beauty filters default to; the goal is the face that was in front of the lens, recovered from softness and noise. For a portrait where you also want to think about lighting, background and skin texture, the sibling page on enhancing a portrait photo goes further into that. And when the photograph is an old print whose faces have gone milky, the photo restorer is the tool built around exactly that job.
Faces are personal data, so the handling is stated in full. Cloud HD is Pro-only and opt-in: your photograph is uploaded to our server for that run and deleted when the run finishes, and in all cases within 24 hours. It is processed in the United States on Cloudflare R2 storage and Modal GPUs, encrypted in transit, and never used to train anything — no face you send teaches any model. The model that ran is named on the result; today it is RestoreFormer++, the model can change, and you can always see which one ran. One enhancement is a single general run, and a failed run is not charged.