A single vision model has been wrong in both directions on this project - it missed real defects and
disputed frames that were correct - so agreement between different model families is worth more than one
confident opinion.
judge_screenshot.py now runs two models by default (openai/gpt-4o-mini + qwen/qwen3.7-flash, both served
on OpenRouter; claude-3.5-haiku is gone, which the first attempt discovered with a 404). combine()
reconciles them: agreement carries the verdict with the mean confidence, disagreement reports 'uncertain'
and keeps both reasons rather than letting the more confident model win. --model still forces a single
judgement, and every line prints which model said what.
comfyui_generate.py generates the same prompt and seed with two checkpoints, scores both by
high-frequency detail (mean absolute Laplacian, which catches blurry or flat results), keeps the loser as
<out>__<model>.png with its sidecar, and moves the winner onto --out. --alt-model restores the single
run.
Tests cover the reconciliation - agreement, disagreement, unknown verdicts, and no verdicts at all - and the
suite is at 103.
Tools/shadow_gap.py measures the vertical pixel gap between an object's lowest pixels and the
top of its shadow, as a fraction of the object's height, and calls it contact or floating. It
classifies a shadow by per-channel scaling - a shadow darkens every channel by the same
proportion, an object does not - which separates the two where an absolute delta cannot. Four
unit tests cover contact, floating, an empty box and the border median.
On the real spawn view it returns 'unknown': the crop contains two surfaces (pale wall behind,
dark road below) and one global background cannot classify both. Recorded as the next step, not
presented as working.
The spawn height is measured and cleared: tracing down from 30 m, the topmost surface under the
player is Z 366 and the capsule's feet are at 369 - 3 cm, the capsule skin. If the character
still reads as floating, the gap is between the rendered pavement and its collision surface, not
the spawn Z, and that is what to measure next.
Two tools that turn character work into something measurable.
- Tools/comfyui_character_blueprint.py generates a four-view sheet (front, side, back,
head) plus a contact sheet from one description, with fixed seeds and a JSON record
of prompts and prompt ids. First sheet: 'GigLine walker', lean pedestrian in a navy
tee, jeans and sneakers.
- Tools/character_match.py compares a rendered character against a sheet: figure
aspect, fill and coverage deltas, mean CIE76 colour distance, dominant-palette
distance, and a side-by-side montage. The deltas stay separate on purpose - a single
similarity score between a photoreal sheet and a low-detail body would be a number
with no meaning.
- comfyui_generate.py grew wait_for_image() so both tools share the polling path
instead of duplicating it.
- 13 unit tests for the colour maths, masks and palette; 95 tool tests pass.
First real result, recorded rather than smoothed over: aspect delta 1.24, fill delta
0.12, mean deltaE 31.9, palette deltaE 25.4. The montage shows the reason - the model
is a nude body and the blueprint wears a tee and jeans, so clothing dominates every
metric. Figure segmentation is explicit (--render-box): auto-detection picked backdrop
blobs on every attempt, and a background plate is the proper fix.
Conflicts resolved by union, not by picking a side:
- AGENTS.md keeps the stronger visual-review rule (README updated with every visual
change; a stale section is a defect).
- render_screenshots.sh keeps the absolute engine path from main and adds
-thirdperson from this branch - both were needed and neither alone worked.
- Screenshots/README.md now lists six shots, retail and third-person both embedded;
the known-issues tables merged, with the animation defect and unproven camera
collision added to main's rows.
- visual_targets.json carries both expectations.
- Architecture docs: ADR-8 (realistic buildings) and ADR-9 (third-person body) both
present and in numeric order; ADR-8 was mis-placed before the Quality Requirements
heading in the combined document by an earlier edit, now fixed.
- Binary conflicts resolved to main's content, since the retail trim only exists
there.
Checks: 82 tool tests, prettier clean, six embeds resolve.
Two parametric cycles authored at 30 fps instead of hand-keyed poses, so the loop
closes by construction: GigLine_Idle (60 f, breathing sway, feet planted) and
GigLine_Walk (30 f, one stride, hip 0.45 rad / knee 0.55 / arm 0.32, arms
counter-phased, torso counter-rotation).
The limb axes were measured, not assumed: a probe rotated upperleg01.L by 0.5 rad
on each axis and read foot.L, which moved -0.022 -> +0.362 in Y on X, twisted on Y
and abducted on Z. Using the wrong axis here would have produced a character
scissoring sideways, which no import error would have caught.
The character export preset now bakes animation. Without bake_anim the exporter
writes a static bind pose and Unreal imports a mesh with no AnimSequence and no
error, so the importer asserts a clip exists and prints VERDICT: PASS/FAIL.
Verified in /Game/Test/Character: Idle 1.967 s and Walk 0.967 s, both AnimSequences,
skeleton intact. Playback on a possessed pawn is not yet verified: the demo pawn has
no visible body, so that check belongs with the third-person camera item.
Models the two retail buildings beside the hero from their own photograph rather
than the generic kit: bronze cornice, mural row, lettering band, stepped Deco
parapet and rounded corner tower on 6800 Lee Drug; burgundy fascia, pilasters,
white-framed windows and banded parapet on 6806-6820. Reference analysis written
before any geometry (Art/blender/retail/ref-analysis.md), every unmeasured
proportion carried into the part report as an assumption.
6822 Budget is deferred, not guessed: no photograph of it exists and the 6.0 x 5.5 m
sliver cannot be read out of the neighbour shot.
Pipeline fixes found on the way:
- render_screenshots.sh called ./Engine/... for the capture step while every
commandlet used the absolute path; set -e killed the script at 127 *after* the
maps were rebuilt, so screenshots silently stayed a day old.
- the facade kit applies to every Frontage_*/North_* actor, so it pasted its own
language over the new modelled facades; it now skips them, and the trim stages as
RetailTrim_*.
- pure bay arithmetic moved to Tools/_common.py so it is unit-testable without
Blender (4 new tests, 82 total pass).
Known gaps recorded in Screenshots/README.md, not hidden: the parapet reads as a
blank band, murals and lettering do not resolve at review distance, the tower is a
plain cylinder, and the review camera crops the east end.
zwitschi