Tie back the Joke Shop curtain bays and add the describe_image vision tool
Full build / Import + headless tests (push) Successful in 39s
Full build / Full desktop build + optional release (push) Skipped

This commit is contained in:
zwitschi committed 2026-11-03 20:58:05 +01:00
1 parent e7d733bc93
commit e4065c2068
6 files changed
+160 -15

No files matched your search

+4
View File
@@ -201,6 +201,10 @@ licence-free zone only.
the `CanopySignatures` label is white, Godot 276/276, `12-joke-shop.png` + the three-up the `CanopySignatures` label is white, Godot 276/276, `12-joke-shop.png` + the three-up
board recaptured. Note: fix 2 (blade sign) is already a circular disc on a roof pole in the board recaptured. Note: fix 2 (blade sign) is already a circular disc on a roof pole in the
code (`37f925c`); the ranking note was stale. code (`37f925c`); the ranking note was stale.
**Fix 5 landed:** the red curtain bays are now seven tied-back drapes (hourglass cinch, dark
recess between bays, scalloped hem) instead of four flat red slabs; vision-judged as
"tied-back/gathered fabric" via `tools/describe_image.py`. Joke Shop map is now 25,828 tris /
518 meshes (under the 150k/700 budget); Godot 277/277, `12-joke-shop.png` + board recaptured.
- [ ] Lighting: `lighting`/`lookdev` pass on sconces, canopy, marquee and street lamps with - [ ] Lighting: `lighting`/`lookdev` pass on sconces, canopy, marquee and street lamps with
day/late captures. (Attenuation is now gated: `test_venue_lights_use_physical_falloff` day/late captures. (Attenuation is now gated: `test_venue_lights_use_physical_falloff`
asserts every `OmniLight3D`/`SpotLight3D` in all three venues uses `attenuation = 1.0`, asserts every `OmniLight3D`/`SpotLight3D` in all three venues uses `attenuation = 1.0`,
Binary file not shown.

Before

Width:  |  Height:  |  Size: 115 KiB

After

Width:  |  Height:  |  Size: 116 KiB

+1 -1
View File
@@ -1,6 +1,6 @@
{ {
"board": "jokeshop-day-board.jpg", "board": "jokeshop-day-board.jpg",
"caption": "Joke Shop frontage, day. Photo: local-only research frame (tmp/research/comedy-store). Concept: photo-anchored (gemini + gpt). Capture: dark stucco wall tile + white signature band.", "caption": "Joke Shop frontage, day. Photo: local-only research frame (tmp/research/comedy-store). Concept: photo-anchored (gemini + gpt). Capture: dark stucco wall tile + white signature band + tied-back curtain bays.",
"photo": "tmp/research/comedy-store/01.jpg", "photo": "tmp/research/comedy-store/01.jpg",
"concept": "docs/screens/concepts/joke_shop/joke_shop_entrance_day-v2.png", "concept": "docs/screens/concepts/joke_shop/joke_shop_entrance_day-v2.png",
"capture": "docs/screenshots/12-joke-shop.png" "capture": "docs/screenshots/12-joke-shop.png"
Binary file not shown.

Before

Width:  |  Height:  |  Size: 553 KiB

After

Width:  |  Height:  |  Size: 540 KiB

+34 -14
View File
@@ -498,20 +498,40 @@ func _build_comedy_store_building() -> void:
step_body.shape_owner_set_transform(step_owner, step_body.shape_owner_set_transform(step_owner,
Transform3D(Basis(), Vector3(door_x, 0.07, front + 0.92))) Transform3D(Basis(), Vector3(door_x, 0.07, front + 0.92)))
root.add_child(step_body) root.add_child(step_body)
# Red curtains hang between the entrance and the signature wall. Three # Red tied-back curtains: seven bays between the entrance and the east end.
# narrow folds and a tieback keep them from reading as flat red slabs. # Each drape is cinched at a tieback into an hourglass, leaving the dark
for x in [-3.2, 1.9, 7.0, 11.2]: # recess visible on either side, and finishes in a scalloped hem. The taper
for fold in [-0.27, 0.0, 0.27]: # plus tone variation makes them read as gathered fabric, not flat slabs.
var fold_color := red.lightened(0.10 if fold > 0.0 else 0.0) for bay in range(7):
_add_box(building, Vector3(0.24, 3.10, 0.18), var x := -3.2 + float(bay) * 2.4
Vector3(x + fold, 2.70, front + 4.56 + absf(fold) * 0.08), # Top drape tapering down to the tieback (widest under the valence).
fold_color, true, null, Vector2.ONE, "cloth") for top in [[4.05, 1.00], [3.55, 0.80], [3.05, 0.60], [2.55, 0.42]]:
_add_box(building, Vector3(0.96, 0.16, 0.22), _add_box(building, Vector3(top[1], 0.52, 0.20),
Vector3(x, 2.13, front + 4.60), Color(0.24, 0.02, 0.03), Vector3(x, top[0], front + 4.56),
false, null, Vector2.ONE, "cloth") red.lightened(0.05), true, null, Vector2.ONE, "cloth")
_add_box(building, Vector3(1.06, 0.52, 0.20), # Tieback cord at the narrowest point.
Vector3(x, 1.34, front + 4.54), Color(0.34, 0.025, 0.035), _add_box(building, Vector3(0.60, 0.14, 0.24), Vector3(x, 2.15, front + 4.58),
false, null, Vector2.ONE, "cloth") Color(0.24, 0.02, 0.03), true, null, Vector2.ONE, "cloth")
# Bottom drape flaring down to the hem.
for bottom in [[1.85, 0.56], [1.48, 0.82]]:
_add_box(building, Vector3(bottom[1], 0.52, 0.20),
Vector3(x, bottom[0], front + 4.56),
red.lightened(0.11), true, null, Vector2.ONE, "cloth")
# Scalloped hem: overlapping shallow spheres along the bottom edge.
for hem_i in range(5):
var hem_x := -0.46 + 0.23 * float(hem_i)
var hem := MeshInstance3D.new()
var hem_mesh := SphereMesh.new()
hem_mesh.radius = 0.25
hem_mesh.height = 0.48
hem_mesh.radial_segments = 8
hem_mesh.rings = 4
hem.mesh = hem_mesh
hem.position = Vector3(x + hem_x, 1.24, front + 4.56)
hem.scale = Vector3(0.62, 0.36, 0.36)
hem.material_override = Materials.flat(red.lightened(0.12), "cloth")
hem.cast_shadow = GeometryInstance3D.SHADOW_CASTING_SETTING_ON
building.add_child(hem)
# Small warm sconces along the black wall. Their bulbs are emissive and each # Small warm sconces along the black wall. Their bulbs are emissive and each
# gets a low-energy practical so the facade is not a black hole after dark. # gets a low-energy practical so the facade is not a black hole after dark.
for x in [-12.4, -6.2, 3.0, 9.4]: for x in [-12.4, -6.2, 3.0, 9.4]:
+121
View File
@@ -0,0 +1,121 @@
#!/usr/bin/env python3
"""Ask an OpenRouter vision model to describe an image (free-form).
Complements the structured judges (`scene_concepts.py realism --judge`,
`screen_designs.py judge`): those return scores/verdicts, this one returns an
open-ended textual observation for exactly the question asked. Used to review
captures and references when the agent harness itself cannot display images.
Usage:
python3 tools/describe_image.py <image> [--question "..."] [--model ...]
python3 tools/describe_image.py <image> <image2> [--question "..."] # compare
Requires OPENROUTER_API_KEY (a gitignored `.env` in the repo root holds it;
source with `set -a && . ./.env && set +a`, or the tool reads `.env` directly).
"""
from __future__ import annotations
import argparse
import base64
import json
import os
import sys
import urllib.error
import urllib.request
from pathlib import Path
ROOT = Path(__file__).resolve().parent.parent
sys.path.insert(0, str(ROOT / "tools"))
import openrouter_judge # noqa: E402
CHAT_URL = "https://openrouter.ai/api/v1/chat/completions"
DEFAULT_QUESTION = (
"Describe this image as an art director reviewing a game screenshot: the "
"subject, composition, materials, lighting, silhouette, and any obvious "
"visual defects or things that read as flat, broken or out of place. Be "
"concrete and specific."
)
def _load_env_file() -> None:
env = ROOT / ".env"
if not env.is_file():
return
for line in env.read_text(encoding="utf-8").splitlines():
if not line or line.startswith("#") or "=" not in line:
continue
key, value = line.split("=", 1)
os.environ.setdefault(key.strip(), value.strip().strip("\"'"))
def _data_url(path: Path) -> str:
return "data:image/png;base64," + base64.b64encode(Path(path).read_bytes()).decode("ascii")
def _call(body: dict, timeout: int = 180) -> dict:
key = os.environ.get("OPENROUTER_API_KEY", "").strip()
if not key:
raise ValueError("OPENROUTER_API_KEY is not set")
req = urllib.request.Request(
CHAT_URL, data=json.dumps(body).encode("utf-8"),
headers={"Authorization": f"Bearer {key}", "Content-Type": "application/json",
"HTTP-Referer": "https://git.allucanget.biz/zwitschi/security-game",
"X-Title": "Gig Line"}, method="POST")
try:
with urllib.request.urlopen(req, timeout=timeout) as resp:
return json.loads(resp.read().decode("utf-8"))
except urllib.error.HTTPError as exc:
raise ValueError(f"OpenRouter HTTP {exc.code}: {exc.read().decode('utf-8', 'replace')[:300]}")
def build_body(model: str, images: list[Path], question: str) -> dict:
content: list[dict] = [{"type": "text", "text": question}]
for path in images:
content.append({"type": "image_url", "image_url": {"url": _data_url(path)}})
return {"model": model, "messages": [{"role": "user", "content": content}],
"temperature": 0}
def main(argv=None) -> int:
ap = argparse.ArgumentParser(description=__doc__.splitlines()[0])
ap.add_argument("images", nargs="+", help="image file(s) to send")
ap.add_argument("--question", default=DEFAULT_QUESTION,
help="question to ask (default: a full art-director description)")
ap.add_argument("--model", default="auto",
help="explicit model, or 'auto' for the cheapest image-capable model")
args = ap.parse_args(argv)
_load_env_file()
images = [Path(p) for p in args.images]
for path in images:
if not path.is_file():
print(f"error: not a file: {path}", file=sys.stderr)
return 2
models = openrouter_judge.resolve_models(args.model)
if len(images) > 1:
args.question = args.question + (
f"\n\nI sent {len(images)} images in order. Compare them and describe the "
"differences, especially anything that changed or regressed.")
try:
model, payload = openrouter_judge.call_with_fallback(
models, lambda m: build_body(m, images, args.question), _call)
except ValueError as exc:
print(f"error: {exc}", file=sys.stderr)
return 1
try:
text = payload["choices"][0]["message"]["content"]
except (KeyError, IndexError, TypeError):
text = "unreadable model response"
if isinstance(text, list):
text = "".join(p.get("text", "") for p in text if isinstance(p, dict))
print(f"[model {model}]")
print(text)
return 0
if __name__ == "__main__":
raise SystemExit(main())