Fixing a face after the clip exists

Two different tools with similar names

Face restoration improves a face that is already the right face but degraded: soft, low resolution, slightly malformed. Models such as GFPGAN and CodeFormer take a face region and reconstruct plausible detail: eyes, teeth, skin texture, edges.

Face replacement puts a different face onto the body in the footage. Different job, different ethics, dealt with below.

Both run per frame on a finished clip. Both are widely available, free, and easy.

When restoration helps

A face at a distance in a 720p generation is perhaps eighty pixels across. The encoder discarded most of its detail and the decoder invented a soft approximation. Restoration is the honest fix for exactly this, and it is the difference between a usable wide shot and an unusable one.

Good cases:

• Faces in the middle distance, soft but correct.
• A close-up where eyes and teeth are mushy but the geometry is right.
• Upscaling a lower-resolution generation where the face is the limiting factor.

Where it fails, and the failure is specific

Temporal flicker. Restoration runs per frame. Each frame’s face is reconstructed independently, so the invented detail differs slightly from frame to frame. The result is a face that boils: skin texture crawling, eye highlights popping, teeth shifting. On a still it looks great; in motion it announces itself.

Mitigations, in order:

• Use a lower restoration strength. CodeFormer’s fidelity weight trades detail against faithfulness to the input; keeping it closer to the input reduces flicker sharply.
• Blend the restored face back at partial opacity over the original. Fifty per cent often looks better than a hundred.
• Use video-aware restoration where available, which considers neighbouring frames.
• Restore fewer frames: on a short shot, restore and check rather than restoring the whole timeline by default.

Identity shift. Aggressive restoration invents a more conventionally attractive, more symmetrical, younger face. It is pulling toward the average of its training data. On a character with distinctive features this quietly replaces them, and on non-white faces the pull toward a training-set average has been a documented and repeatedly criticised problem. Check the output against your character sheet rather than against your impression.

Face replacement, and the line

The technical description is short: detect the face, align it, swap in a face generated or transferred from a source, blend the edges, run per frame.

Legitimate uses exist and are real. Holding your own trained character across shots where the generation drifted. Replacing a background actor’s face with a synthetic one to avoid needing their release. Fixing one shot in a sequence where everything else matched.

The illegitimate use is the one this technology is famous for, and the mechanics are identical. There is no technical difference between the operation that rescues your shot and the operation that puts a real person into a video they were never in.

So the line is not technical, it is about permission and disclosure:

• A real person’s face requires that person’s written, specific permission. Consent to be filmed is not consent to be synthesised.
• The output requires a visible label, for reasons dealt with properly in the final module.
• Many jurisdictions now legislate here specifically, and the direction of travel is one way. Non-consensual intimate synthetic imagery is criminal in a growing number of places, and personality-rights actions over synthetic likeness have succeeded in several courts since 2023.

If you find yourself reasoning about why a particular case is an exception, that is the signal to stop and ask the person.

The free path

GFPGAN, CodeFormer and the current successors are free and open, available as Hugging Face Spaces for browser use and as ComfyUI nodes locally. FFmpeg extracts frames and reassembles them. Resolve’s free tier blends a restored sequence back over the original with an opacity slider.

None of this needs a paid tool, which is precisely why the ethical question cannot be left to whoever can afford the software.

Today: restore one face in a soft wide shot at low strength, blend it back at fifty per cent, and watch the result in motion rather than as a frame.