Updated
Case evidence: Read the source breakdown report · View case frames and shot timings
Based on a published tool report. Production classifications are analytical judgments; suggested steps are not the original project or a verified recreation.
What a video face swap actually does
A swap pipeline does three things per frame: it detects the destination face and its landmarks (jaw, brows, mouth line), warps the source face to match that geometry, and blends the result back with edge smoothing and color matching. Older swaps needed a training run per face; modern one-shot models swap from a single photo, which is why the tools multiplied. Quality depends less on the model than on the footage: frontal, evenly lit, steady faces swap convincingly; profiles, occlusions (hands, hair, microphones) and fast motion are where seams survive. A swap differs from a talking photo (which animates one still) and from a full avatar (which generates the whole person) — it keeps the target’s performance and replaces only the face.
The tells, frame by frame
Work through the clip one pause at a time and look in five places. Edges: the jaw-hairline-ear rim, where warped skin meets original skin — look for a tone or grain step, especially when the head turns. Teeth and ears: high-frequency detail the generator rebuilds worst; count the teeth, watch inner-ear structure. Lighting: swapped skin often ignores the scene’s key light direction; squint and compare cheek highlights to the background. Blink and breath rhythm: swapped faces blink on model cadence, not the body’s. Sync and identity drift: over long clips the face slowly drifts back toward the source’s geometry — compare the first and last frames side by side. Any single tell is a reason to check further, never a verdict.
The tool field
Searches mostly meet three groups. One-click web and app swappers (Supawork, Remaker and dozens of clones) trade control for convenience — credit-priced, watermark-free tiers, terms that vary on what happens to your uploads. Creator-grade tools (roop-class open-source forks, DeepFaceLab lineage) run locally on your GPU and give control over every stage, at the cost of setup skill. Platform “effects” (some social apps’ built-in face filters) are swaps constrained to live video. Before uploading anyone’s face anywhere, read where the file goes: a local run is the only route where nothing leaves your machine.
How to run a local check with rea-video
When a clip matters — a claim, a payment, a reputation — run the open-source rea-video workflow on it. It decomposes the video into a shot table, flags any shot whose face behavior, boundary texture or lighting pattern reads as manipulated, and attaches the evidence frames and a confidence estimate to each row, so a claim like “this confession is a swap” becomes checkable: here are the frames, here is the seam, here is what compression may already have erased. It runs on your machine, which matters when the clip itself is sensitive. The report will also tell you when it cannot tell — low-confidence rows are listed, not hidden.
The consent line
Technical quality is not permission. Swapping a face without the source person’s consent — onto explicit, criminal or merely embarrassing footage — is unlawful in a growing list of jurisdictions regardless of how convincing the blend is, and platform synthetic-media rules increasingly require labeling. The defensible uses are narrow: your own face on your own footage, consenting collaborators, satire with clearly disclosed manipulation, and VFX with releases. If you cannot name the consent, you do not have it.
Pertanyaan yang sering diajukan
Are Supawork and Remaker good face swap tools?
They are the most-searched of the current one-click swappers: fast, credit-priced, no GPU needed. That convenience trades away control — your photos and footage go to their servers, terms vary, and quality on profiles and fast motion drops. For sensitive material, only a local pipeline keeps files on your machine.
How can I tell if a video is a face swap?
Pause through the clip and check five places: the jaw-hairline-ear edge for tone or grain steps, teeth and inner ears for rebuilt detail, cheek highlights against the scene’s light direction, blink cadence against the body’s breathing, and first-vs-last frame identity drift. Each tell is a reason to check further — for a verdict-grade answer, run a local breakdown and read the evidence rows.
Is face swapping legal?
Swapping your own face onto your own footage is generally fine. Swapping someone else’s face without consent — especially onto explicit, criminal or embarrassing material — is unlawful in a growing number of jurisdictions regardless of quality, and platforms increasingly require synthetic-media labels. If you cannot say where the consent came from, assume there is none.
Can rea-video detect face swaps?
rea-video does not output a percentage score. It produces a per-shot account with evidence: which shots read as manipulated, which frames show the seam, what compression may have erased, and a confidence estimate per row — so you can verify the reasoning yourself instead of trusting a number. It runs locally, which matters when the clip is sensitive.
A reusable analysis template
Clip source: original file or re-upload? (compression!) Face size in frame: large / medium / small Edges: jaw / hairline / ears — tone or grain step? Teeth count & inner ear: consistent? Lighting: cheek highlight vs background key light Blinks: cadence natural? breathing matches body? First vs last frame: identity drift? Check: local rea-video breakdown → evidence rows + confidence Claim status: supported / unsupported / inconclusive Consent: named and documented?
Practice with a real report
Practice reading face evidence on the music-video case: its talking-avatar shots carry per-shot face rows — compare how generated faces behave at boundaries versus filmed ones.
Open the case and download its report →Continue your workflow
- Video to prompt: translate observations into instructions
- AI video prompt library: 20 copyable prompts by scenario
- Which AI video model made this? Kling 3.0, Seedance 2.5, Runway Gen-4.5 and how to tell
- AI talking photo: the 2026 field guide, tools and prompts
- How to make a talking photo: step by step (2026)
- Reverse any video: shots, prompts and a remake plan
- Get started in Claude Code or Codex