Localize, Don't Beautify

LaTeX source and analysis artifacts for "Localize, Don't Beautify: Client-Side Control of Image-Editing APIs for Cosmetic Surgery Previews" (IEEE conference format).

This pilot tests whether a landmark-derived, client-side composite can prevent image-editing APIs from changing pixels outside a requested facial region. It can: the composite enforces preservation by construction, and target-region pixel change is largely retained. The study does not establish clinical plausibility or predict surgical outcomes; it contains 15 main-analysis faces, one output per cell, and no surgeon ratings.

Repository scope

The repository contains the canonical score table, aggregation code, generated tables/plots, qualitative figures, and paper source:

experiment run records (separate, non-public repository)
        |
        v
scripts/aggregate.py  --> data/canonical_rows.csv
                     --> tables/main_results.tex
                     --> tables/setup_models.tex
                     --> tables/numbers.tex
scripts/make_figures.py     --> figures/generated/*.pdf
scripts/make_qualitative.py --> figures/generated/*.png

The committed canonical CSV is sufficient to audit the paper's quantitative summaries and regenerate quantitative plots. It is not sufficient for end-to-end reproduction of API generation, masking, alignment, scoring, or qualitative image selection. PLASTYVUE_POC must point to the separate experiment repository to rebuild from raw run records.

Building

You need a TeX distribution with latexmk and uv.

make numbers    # requires the separate experiment repository
make figures    # quantitative plots use the committed CSV; qualitative grids need source images
make pdf        # latexmk -> main.pdf
make check      # verify citation parity

Or run make pdf with the committed generated artifacts.

Identifiable source images

The qualitative figures contain publicly accessible before/after photographs of real people, used with permission from the source clinics.