Original document
The original paper
The longer explanation of SkinDeep’s preference-learning approach.
The paper describes learning from ratings and using the result to guide generated images. It also includes ideas for other applications.
Some claims in the older paper go beyond what has been tested. The current corrections are below.
Corrections and technical notes
Historical claims and forecasts are preserved in the archive. Use the maintained method and measured results for current technical guidance.
Read the original
Technical errata
- Inversion: a score usually maps to many latent inputs. Specify an optimization objective, bounds, and feasibility.
- Classifier: the browser demo uses logistic regression; SVMs and deeper feature-space networks are different alternatives.
- Performance: millisecond training is a scoped measurement, not end-to-end generation latency. Old multiplier claims lack attached reproducible runs.
- Multiple modes: randomized search does not give a linear classifier disconnected preference regions.
- Privacy and validity: embeddings and negative examples do not establish universal guarantees.
- Transfer: generators and domains need compatible representations or learned alignment.