image-replicator

image-replicator

Purpose

After the blockout is plausible, this sub-skill drives the detail pass. Each iteration produces a render, runs the scorer, and returns a concrete decision: which gate is weakest, what model change closes it, whether to continue or stop.

Inputs

• The .kcad.ts file with a clean blockout.
• The reference photo path.
• Current scorer output (or "not yet run" on the first pass).

The iteration loop

Step 1 — Render

Always render all four canonical views at the correct resolution:

kernelcad render build.kcad.ts \
  --width 1920 --height 1080 \
  --views front,right,top,iso \
  -o /tmp/render.png

Read the PNG back. Filenames are not evidence — you must observe the image.

Step 2 — Score

kernelcad score build.kcad.ts \
  --reference ./reference.jpg \
  --view front \
  --json

The scorer returns a JSON object with per-gate values:

{
  "silhouette": 0.71,
  "ssim": 0.28,
  "phash": 0.62,
  "overall": 0.54
}

Score all views that have a reference. If the task harness provides a harness.ts, run it directly — it wraps scoreAgainstReference with the correct thresholds:

npx ts-node eval/tasks/taskNN/harness.ts

Step 3 — Diagnose which gate to chase next

Read the per-gate numbers and decide:

Gate	Low score means
silhouette	The 2D silhouette outline does not match: wrong overall shape, missing brow bumps, wrong arm angles, wrong aspect ratio. Fix: adjust params, fix the path outline.
ssim	The rendered image is structurally different from the reference: wrong depth, wrong material (flat vs. glossy), missing lens geometry, wrong body proportion. Fix: apply PBR material (Rule 6), add depth (Rule 3), fix booleans.
phash	Perceptual hash mismatch: high-level appearance differs. Usually a proxy for "the object reads as a different product category." Fix: re-read the brief, re-check primary mass arrangement.

Chase the lowest gate first. A 0.20 silhouette score means the silhouette is wrong — fixing the SSIM first will not help because the silhouette mismatch dominates overall score.