FPEM and LiveBeauty: A Candidate Baseline, Not a Ready-Made Verdict
What does an available specialist model add beyond asking a general multimodal model for a rating?
A baseline must be reproducible
An available training and test workflow makes it possible to examine preprocessing, inputs and model outputs rather than infer the method from a polished demo. We would freeze those choices, record the checkpoint version and test the same independent images used for our existing baselines. Merely downloading weights would not complete a comparison.
The domain can change the answer
Live-stream frames, retouching and ordinary user portraits can differ in pose, image processing and presentation. A model’s original use case therefore matters when interpreting its output on a front-facing selfie. We would measure performance and failure cases on the photos our tools actually accept instead of assuming cross-domain transfer.
Permission is a separate deployment gate
The repository identifies a code license and describes accessing model and dataset material. Code availability does not by itself settle every right relevant to a commercial service. Before an experiment involving downloaded material, we need to verify the applicable checkpoint and dataset terms and the relevant portrait permissions. Our research article does not make those permissions exist.
Our adoption status
FPEM is our first candidate for an offline specialist-model evaluation because its project exposes a test workflow. It is not currently serving our users. Our public report therefore describes measured geometry and AI photo observations, with human-rating prediction listed as a validation objective rather than a deployed feature.
What we have adopted
Candidate for an offline evaluation. A code license alone does not settle checkpoint, training-data or portrait permissions. We have not deployed its weights.
Our public methodology documents the deployed evidence layers and proposed validation protocol. For a concrete example, inspect the complete sample report. The example is illustrative, not an experimental validation result.
Continue reading
Validating Face Analysis: Independent People, Independent Ratings
What an independent human-rating evaluation must establish before a face-analysis product claims validated predictive performance.
TransFBP: What Whole-Face Attention Can Teach a Face Analyzer
A research note on TransFBP, cross-attention, whole-face relationships and the difference between an attention map and a causal explanation.