"""s3 — 실외 체인: 뿌리 스케치 마킹(O1) → 배경 plate(O2) → 임시샷 → A/B/C → 최종. fp 미사용 — 뿌리(조감 스케치)가 유일한 공간 control. 사용: .venv/bin/python s3_outdoor.py --root A|B [--only mark|plate|paths] """ import argparse import json import sys from pathlib import Path sys.path.insert(0, str(Path(__file__).parent)) import forest_lib as F # noqa: E402 import forest_chain as C # noqa: E402 OUTDIR = F.OUT / "outdoor" # --variant v2 시 main()에서 교체 MARK_PROMPT = ( "The attached image is a TOP-DOWN aerial site sketch with circled capital" " letters marking site elements (and possibly a legend panel on the right)." " Keep the sketch EXACTLY as it is — same geometry, same circled letters," " same legend, do not redraw or restyle anything.\n" "ADD exactly these annotations for ONE camera setup:\n" "- ONE green camera icon with a green triangular view-CONE showing where the" " camera stands and what it looks at, following this staging description:\n" "{cam}\n" "{figures}" "Add NOTHING else — no new structures, no new text other than the new" " capital letters in their small circles." ) def figures_clause(s) -> str: st = s.get("staging") or {} cons = ((st.get("frame_spatial_contract") or {}).get("constraints") or []) n = sum(1 for c in cons if isinstance(c, dict) and c.get("target_kind") == "character") if n <= 0: n = len(st.get("character_angles") or []) if n <= 0: return "- This shot has NO people — add no lettered person circles.\n" letters = " ".join(f"({chr(80 + i)})" for i in range(n)) # P, Q, R... (요소 A~와 구분) return (f"- {n} red circle(s) with capital letters {letters} inside the" f" view-cone where the staging places the people:\n" + "\n".join(" * " + json.dumps(ca, ensure_ascii=False) for ca in (st.get("character_angles") or [])) + "\n") def o1_mark(recon, sk, root_png): s = recon["shots"][sk] cam = str((s.get("staging") or {}).get("camera_direction") or "") prompt = MARK_PROMPT.format(cam=cam, figures=figures_clause(s)) return F.img_gpt(f"{sk}_rootmark", prompt, refs=[root_png], out_path=OUTDIR / sk / "root_marked.png") def o2_plate(recon, sk, root_plan, extra: str = ""): s = recon["shots"][sk] marked = OUTDIR / sk / "root_marked.png" locs = [m for m in recon["members"] if m["loc_id"] in s["loc_ids"]] loc_lines = "\n".join(f"- {m.get('label')}: {m.get('summary')}" for m in locs) prompt = "\n".join([ "The attached image is a TOP-DOWN aerial site sketch of one real place,", "with a green camera icon and view-cone marking ONE camera position and", "its viewing direction. Circled letters mark site elements; red lettered", "circles (if any) mark where people stand — but render NO people.", "", "Draw the PHOTOREALISTIC outdoor view that THAT camera sees from its", "marked position at eye level, looking along the view-cone: the", "structures, ground surfaces, openings and approaches the sketch places", "in front of that camera, at real-world human scale. This is an EMPTY", "background plate: absolutely no people, no text, no letters, no", "circles, no arrows, no sketch lines — pure photographic realism.", "", "WHAT THIS PLACE LOOKS LIKE (production data):", loc_lines, "", "Site orientation notes:", str(root_plan.get("sketch_notes", "")), "", C.staging_lines(s), ] + ([extra] if extra else [])) return F.img_gpt(f"{sk}_plate{'_v2' if extra else ''}", prompt, refs=[marked], out_path=OUTDIR / sk / "plate.png") def run_shot(recon, sk, mannequin_preview: bool, v2: bool = False): s = recon["shots"][sk] d = OUTDIR / sk env = d / "plate.png" control = d / "root_marked.png" passports = C.passports_for_shot(recon, s) tagv = "_v2" if v2 else "" contracts = ([C.people_contract(s), C.MARKER_NOT_PEOPLE, C.EYE_LEVEL_CONTRACT] if v2 else None) temp = C.temp_shot(f"{sk}_temp{tagv}", s, env, control, d / "temp.png", contracts=contracts) # 경로 A: 직행 C.final_shot(f"{sk}_final_A{tagv}", s, env, passports, d / "final_A.png", control_png=control, contracts=contracts) # 경로 B: VLM 구도 → 배경 재생성 → 최종 comp = C.vlm_composition(sk + tagv, temp) bg_b = C.bg_regen(f"{sk}_bgB{tagv}", s, env, d / "bg_B.png", comp=comp, contracts=contracts) C.final_shot(f"{sk}_final_B{tagv}", s, bg_b, passports, d / "final_B.png", contracts=contracts) # 경로 C: 윤곽 → 배경 재생성 → 최종 cont = C.contour(f"{sk}_contour{tagv}", temp, d / "contour.png") bg_c = C.bg_regen(f"{sk}_bgC{tagv}", s, env, d / "bg_C.png", contour_png=cont, contracts=contracts) C.final_shot(f"{sk}_final_C{tagv}", s, bg_c, passports, d / "final_C.png", contracts=contracts) if mannequin_preview: bg_cm = C.bg_regen(f"{sk}_bgC_m{tagv}", s, env, d / "bg_C_m.png", contour_png=cont, mannequin=True, contracts=contracts) C.final_shot(f"{sk}_final_C_m{tagv}", s, bg_cm, passports, d / "final_C_m.png", contracts=contracts) def main(): global OUTDIR ap = argparse.ArgumentParser() ap.add_argument("--root", required=True, choices=["A", "B"], help="s1 육안 선정 root 변형") ap.add_argument("--only", default="all", choices=["all", "mark", "plate", "paths"]) ap.add_argument("--variant", default="v1", choices=["v1", "v2"], help="v2 = 인원/마커/eye-level 계약 추가 (1.5라운드)") args = ap.parse_args() v2 = args.variant == "v2" if v2: OUTDIR = F.OUT / "outdoor_v2" recon = F.load_recon() root_plan = F.load_plan("root_plan") root_png = F.OUT / "root" / f"root_{args.root}.png" if not root_png.exists(): raise SystemExit(f"root image missing: {root_png}") subset = F.load_plan("subset")["outdoor"] for i, sk in enumerate(subset): if v2: # 마킹은 v1 산출 재사용 (동일 입력 — 재생성 불필요) src = F.OUT / "outdoor" / sk / "root_marked.png" dst = OUTDIR / sk / "root_marked.png" if src.exists() and not dst.exists(): dst.parent.mkdir(parents=True, exist_ok=True) import shutil shutil.copyfile(src, dst) if args.only in ("all", "mark") and not v2: o1_mark(recon, sk, root_png) if args.only in ("all", "plate"): o2_plate(recon, sk, root_plan, extra=(C.EYE_LEVEL_CONTRACT if v2 else "")) if args.only in ("all", "paths"): run_shot(recon, sk, mannequin_preview=(i == 0), v2=v2) F.runlog({"kind": "stage", "stage": "s3_outdoor", "done": args.only, "root": args.root, "variant": args.variant, "subset": subset}) if __name__ == "__main__": main()