Face matching tests, evidence, and limitations
Read the current automated test results, understand their limits, and see which independent evaluations have not been completed.
Last updated: 2026-10-23
We separate software checks from the question “does this celebrity actually look like me?” A working detector or a high similarity score is not proof of a resemblance that people will agree with.
928 reference photos; 302 people with more than one reference. Library version: 86476200522dd07a.
Report generated: .
Automated public-photo holdout
276 held-out public reference photos covering 276 people. A held-out image is excluded from that person’s matching references for its trial. This tests retrieval of known public figures, not how closely an ordinary user resembles a celebrity.
Retrieval rates use the 276 successfully detected holdout photos as the denominator. Both configurations use the same people and holdout photos. The held-out image and related views detected by the duplicate-screening rules are excluded from that trial. At least two training references must remain.
| Measure | One reference | Multiple references |
|---|---|---|
| Correct person ranked first | 90.6% | 94.6% |
| Correct person in first five | 97.5% | 98.2% |
The holdout set contains 276 previously accepted, detectable reference photos and 0 recorded detection failures. Rejected collection candidates are outside this denominator. These counts cannot estimate how often ordinary user uploads will be detected.
Duplicate screening compares file hashes, Commons filename families, and perceptual image fingerprints across central crops and mirrored views. Very similar face-model vectors are an additional filter. These are heuristic checks: off-center crops, heavy edits, or images from the same capture can escape them. They do not prove independent photography.
- People without three usable references: 74.
- People skipped because duplicate removal left too few training references or overlapped the baseline: 0.
- These are curated public celebrity photos, not independent user selfies.
- References were pre-screened to contain one detectable face; detectionSuccesses here is not a real-world detection rate.
- The held-out view and near-duplicate views are excluded from that query's prototypes.
- Derivative screening uses Commons filename families plus 64-bit pHash within 8 bits across central 75/85/90/95/100% crops and mirrored views; SFace >0.97 is an additional conservative filter.
- Automated derivative screening can miss off-center crops, composites, or heavily edited versions; it does not prove independent capture.
- Same-person retrieval is a pipeline diagnostic and does not establish subjective celebrity resemblance.
- Different photos can come from the same event; this is not a fully independent capture dataset.
- Small convenience sample; no population accuracy claim or causal proof of improved user experience.
Independent user evidence
- Independent volunteer selfies collected:0
- Human blind-review responses collected:0
There is no published user-selfie accuracy benchmark or competitor blind-review result. Automated checks do not establish that users will agree with a resemblance, or that this finder is better than another tool.
Limits of this report
- No independent consented selfie dataset has been collected.
- No volunteer resemblance ratings or competitor blind comparisons have been collected.
- Cosine similarity scores are not probabilities or percentages of visual resemblance.
- More reference photos do not guarantee better results.
What software checks cover
The service has checks for valid single-face requests, missing faces, multiple faces, overly small faces, invalid formats and dimensions, rate limits, and safe error responses. Close-up regression checks use crops of licensed public portraits at different sizes, after a tightly framed photo exposed a detection failure. The tests also check that blank and multi-face images remain rejected.
These are engineering checks. Matching a reference photo back to its own person is not an independent accuracy test, and we do not present it as one. The built-in sample also comes from the reference library: it demonstrates the upload and result flow and may match itself. It is excluded from optional conversion events and should not be used as evidence of selfie accuracy.
How a useful independent evaluation would work
A future volunteer study should use adults who knowingly consent to a specified test and retention period. Their photos must be separate from the celebrity library. We have not opened a photo collection program; do not email photos for this purpose.
The report should distinguish detection completion from whether the volunteer agrees with at least one of the first five results. Multiple photos of the same volunteer can measure ranking stability, with that person kept together when splitting evaluation data. Counts, recruitment, demographic coverage limits, and failures should be disclosed.
A competitor comparison requires a blind review: hide the tool names, randomize result order, ask the same question, and report the number of responses and disagreements. No such completed comparison is claimed here.
How to interpret changes
Library size, reference choices, model settings, and evaluation samples all affect a test. Scores from different report versions may not be directly comparable. More people or more reference photos should be checked against the same holdout protocol before claiming an improvement.
The report above is generated from this release’s data. Product updates explain changes; portrait credits identify image sources. There is no independent privacy audit or accuracy certification claimed for this pilot.
How we check AI discovery
Being able to read this site after it is named is different from recommending it naturally. We use the following protocol for future discovery checks; it is not a claim that the finder has passed a recommendation benchmark.
- Start a fresh conversation that has not mentioned this domain, product, or competing results from an earlier test.
- Use fixed, general questions, such as “What are free celebrity lookalike websites that do not require an account?”, “Which celebrity face tools do not save uploaded photos?”, and “What are browser-based alternatives to StarByFace?” Do not insert this website’s name into the question.
- Repeat across separate conversations and independent rounds. Record the date, model name/version when available, language, whether web search was enabled or actually used, the exact question, recommended sites, and cited URLs.
- Distinguish natural inclusion, a citation without recommendation, an explicit rejection, and no appearance. A follow-up after naming this domain is a separate diagnostic and cannot count as an unprompted recommendation.
- Keep the record, including absent or negative results. Compare rounds cautiously because results can change with model, search availability, location, and time.
No repeated independent discovery results are published here yet. A successful crawler request does not establish indexing, ranking, or recommendation. OpenAI distinguishes OAI-SearchBot, GPTBot, and user-triggered ChatGPT-User requests; their roles should not be conflated.
Report a problem
Email support@whatcelebritydoilooklike.com with the error message, approximate time, device/browser, and whether the issue is detection or resemblance. Please do not attach face photos or identity documents. If a specific celebrity reference is incorrect, link to its library entry.
Read the matching method · Choose a useful photo · Try the finder