widget2code-bench

Benchmark evaluation for widget code generation — 12 quality metrics across layout, legibility, perceptual, style, and geometry.

Installation

# 1. Install PyTorch with CUDA support first (skip if CPU-only)
pip install torch torchvision --index-url https://download.pytorch.org/whl/cu126

# 2. Install widget2code-bench
pip install widget2code-bench

Note: PyPI only ships CPU-only PyTorch. To use --cuda, you must install PyTorch from the official index before installing this package.

Usage

Single image mode

Evaluate one GT-prediction pair. Prints JSON results to stdout, no files saved.

widget2code-bench \
  --gt_image /path/to/gt.png \
  --pred_image /path/to/pred.png \
  --cuda

Batch mode

Evaluate all matched pairs in directories.

widget2code-bench \
  --gt_dir /path/to/GT \ # /shared/zhixiang_team/widget_research/Comparison/GT
  --pred_dir /path/to/predictions \
  --pred_name output.png \
  --cuda

Directory Structure (batch mode)

• GT dir: flat image files with 4-digit IDs in filenames (e.g. gt_0001.png)
• Pred dir: subfolders with 4-digit IDs in names, each containing --pred_name file

gt_dir/                     pred_dir/
  gt_0001.png                 image_0001/
  gt_0002.png                   output.png
  ...                         image_0002/
                                output.png

Options

Output (batch mode)

Per-sample outputs

Every matched pair writes one evaluation.json plus (by default) a full per-metric visualization set into its sample folder:

<pred_dir>/
  image_0001/
    output.png
    evaluation/
      evaluation.json                 # 12 metrics
      viz/
        MarginAsymmetry.png
        ContentAspectDiff.png
        AreaRatioDiff.png
        TextJaccard.png
        ContrastDiff.png
        ContrastLocalDiff.png
        PaletteDistance.png
        Vibrancy.png
        PolarityConsistency.png
        ssim.png
        lp.png
        geo_score.png

Each viz PNG shows left/middle = GT/Pred intermediates and right = formula + intermediate values + final score, so you can see exactly how the metric was computed.

Pass --minimal to skip the viz/ directory (much faster, ~10x less disk).

Missing-prediction handling

The evaluator always produces all four fill modes. When a GT image has no matching prediction:

• Existing subfolder, pred missing → fill results go in the same folder's evaluation/
• No subfolder at all → evaluator creates pred_dir/fill_<id>/evaluation/

In either case it writes:

evaluation/
  evaluation_black.json   # GT vs all-black image
  evaluation_white.json   # GT vs all-white image

zero fill isn't a per-sample file — it's a worst-case contribution (LPIPS = 1.0, others = 0) used only when aggregating the combined summary.

Aggregate outputs (.analysis/)

<pred_dir>/.analysis/
  metrics_stats.json                 # per-metric quartiles/mean/std over matched pairs
  metrics.xlsx                       # 4-row combined summary (raw/black/white/zero)
  raw/<run>-raw-<ver>.xlsx           # single-row summary per mode
  black/<run>-black-<ver>.xlsx
  white/<run>-white-<ver>.xlsx
  zero/<run>-zero-<ver>.xlsx

All numeric values are rounded to 2 decimals. Combined metrics.xlsx has a two-level header grouping metrics by category (Layout / Legibility / Style / Perceptual / Geometry) plus SuccessRate (ratio, count). Per-mode xlsx uses flat single-level headers.

All metrics are higher-is-better except lp (LPIPS), which is a distance (lower-is-better).

License

Apache-2.0