Skip to content

Footwear QC · measured, not marketed

How good are our shoe renders, really?

Every figure below is measured on a real benchmark run of shoe input/output pairs and shown with its methodology — sample size, library, pipeline version, and the date it was measured. No figure can appear without them. We report the hard presets honestly: a recolor's colour drift and a try-on's stance drift sit right next to the easy wins.

The pipeline that produced these numbers

  1. 1

    Input shoe image

    A real product or on-model shoe photo.

  2. 2

    oaktree/image-edit

    Single Oaktree edit pass. No -pro variant.

  3. 3

    Output render

    The generated shoe image.

  4. 4

    QC measurement

    colour-science + MediaPipe + pixel-spec.

Model

oaktree/image-edit

Cost / image

$0.04

Inputs → outputs

24 → 24

Total run cost

$0.96

measured 2026-06-02 · pipeline v1.0

Objective metrics

computed

Each figure is computed by a named library against a stated threshold. Hard presets first — their lower numbers are the point, not a thing to hide.

Catalog Variant Generator (Recolor)

oaktree/image-edit · $0.04/image

hardest preset

Hardest preset for colour: a recolor must hit a target hue while holding material, stitching and specular highlights. ΔE is the metric that bites here — we report it honestly.

Color fidelity (ΔE 2000)

target ≤ 10 ΔE to target

7.75 ΔE pass
n=8 · colour-science 0.4.7 · pipeline v1.0 · measured 2026-06-02

Background compliance

target ≥ 95% PDP pass

61.8% below target
n=8 · opencv/numpy 4.13.0 · pipeline v1.0 · measured 2026-06-02

Auto-QC first-pass rate

target ≥ 85% first pass

50.0% below target
n=8 · pixfino-qc-gate n/a · pipeline v1.0 · measured 2026-06-02

Virtual Try-On (Footwear)

oaktree/image-edit · $0.04/image

hard preset

On-model shoe try-on is hard: the render must keep the model's identity and foot stance while swapping the shoe. Identity and pose drift are the honest pressure points.

Face preservation

target ≥ 0.99 landmark cosine

0.999 pass
n=6 · mediapipe 0.10.35 · pipeline v1.0 · measured 2026-06-02

Pose / stance preservation

target ≤ 0.15 stance drift (torso-norm)

0.023 pass
n=6 · mediapipe 0.10.35 · pipeline v1.0 · measured 2026-06-02

Color fidelity (ΔE 2000)

target ≤ 2.0 ΔE (imperceptible)

PLACEHOLDER · run harness n=0 · date TBD

Auto-QC first-pass rate

target ≥ 85% first pass

100.0% pass
n=6 · pixfino-qc-gate n/a · pipeline v1.0 · measured 2026-06-02

White-Background Standardizer

oaktree/image-edit · $0.04/image

high pass-rate baseline

The easiest of the three and our highest pass rate — included as an honest contrast so the harder presets read in context, not as a wall of identical 99%s.

Background compliance

target ≥ 95% PDP pass

98.3% pass
n=10 · opencv/numpy 4.13.0 · pipeline v1.0 · measured 2026-06-02

Color fidelity (ΔE 2000)

target ≤ 2.0 ΔE (imperceptible)

PLACEHOLDER · run harness n=0 · date TBD

Auto-QC first-pass rate

target ≥ 85% first pass

80.0% below target
n=10 · pixfino-qc-gate n/a · pipeline v1.0 · measured 2026-06-02

Per-render certificates

Every render carries its own QC certificate with a shareable link — its actual ΔE, white-bg compliance, identity and stance numbers, not an average.

Subjective notes

not a metric

These are human impressions from reviewing the set — deliberately kept out of the numbers above. "Looks good" is not a measurement. Recolors read clean on matte materials and get harder on glossy patent finishes; on-model try-ons are most convincing in neutral standing poses. Treat this as context, not as a score.

Theme
Type scale
A A