# Enhance experiment assets

The five PNG files are unmodified copies of the saved September 25, 2026 local model experiment. They are not generated marketing illustrations or output from the October 5 online checks.

| Published file | Original experiment file |
| --- | --- |
| comparison-compressed.png | comparison-compressed.png |
| clean-source.png | clean_video_source_bicubic2x.png |
| clean-x2plus.png | clean_video_RealESRGAN_x2plus.png |
| bruce-lee-source.png | old_video_source_bicubic2x.png |
| bruce-lee-x2plus.png | old_video_RealESRGAN_x2plus.png |
| local-model-measurements-2026-09-25.json | measurements.json |

The experiment used 512 × 512 video-frame inputs, an RTX 4060 Laptop GPU and ONNX Runtime DirectML. GPU medians are from three runs after warm-up, excluding model loading, CPU preprocessing, decoding and encoding. The source panels are bicubic enlargements, not native-resolution AI outputs. NAFNet produces 1× output and is enlarged only for display.

Model exports differed: SPAN used dynamic FP32; x2plus used 288 × 288 FP16 tiles with 16-pixel overlap; RealPLKSR MSSIM used fixed 512 × 512 FP32; PureScale GAN used fixed 512 × 512 FP16. These settings describe that historical experiment, not today's online deployment or an equal-precision architecture benchmark.

Rebuild SVG charts with `node scripts/build-enhance-benchmark-charts.mjs`. Local charts use the archived JSON here; online charts use `../image-enhancement-2026-10-05.json`. No quality score, CPU baseline, competitor result or universal latency claim is inferred from these measurements.