DisFace3DNet: Explaining a Prediction Without Claiming a Cause

Separate the photograph from the face

The practical design lesson is to distinguish what appears in the facial outline from how the image is lit, framed and posed. In our method, visible surface appearance and photographic presentation are separate dimensions. A shadow or a soft image must not be turned into a claim about skin health or permanent facial quality.

A component term belongs to a particular model

A model can allocate numerical contributions according to its architecture and training constraints. Those contributions describe its computation. They do not prove that changing hair in the real world will change a human panel’s opinion by the same amount. That stronger claim would require a separate experimental design.

Our report is an observation system

Current Get Face Report narratives are grounded in the submitted photo and supplied measurements. We do not run DisFace3DNet’s 3D component decomposition, export its numerical contributions or treat a VLM explanation as a substitute. Our labels distinguish measured evidence, a visual observation and a suggested styling comparison.

How a reader can check an explanation

A useful report should let you locate the feature or image condition it describes. It should name an alternative interpretation when one matters, such as a head turn affecting the visible jaw. A recommendation should specify what to keep constant, what to change and how to decide which version you prefer.

What we have adopted

Informs the separation of feature appearance and photo context. Our current report does not perform this 3D decomposition or compute its component contributions.

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