"""실험 — interior 도면을 reference로 4방향(N/E/S/W) 사진 멀티턴 생성. 사용: python scripts/experiment_360_interior.py \ --plan-png --out-dir [--run-dir ] run-dir 지정 시 step1_spatial.json에서 environment_canon 추출하여 prompt에 자동 주입. """ from __future__ import annotations import argparse import base64 import json import logging import os import sys from pathlib import Path from typing import Optional BACKEND = Path(__file__).resolve().parent.parent sys.path.insert(0, str(BACKEND)) from dotenv import load_dotenv # noqa: E402 load_dotenv(BACKEND / ".env") from openai import OpenAI # noqa: E402 logging.basicConfig(level=logging.INFO, format="%(asctime)s %(levelname)s %(message)s") logger = logging.getLogger("360") def compact_canon_text(canon: dict) -> str: parts = [] b = canon.get("building") or {} if b: bp = [] for k in ("stories", "primary_material", "exterior_stairs", "rooftop_features", "window_pattern", "weathering"): v = b.get(k) if v: if isinstance(v, list): v = ", ".join(map(str, v)) bp.append(f"{k}={v}") if b.get("color_palette"): bp.append("colors=" + ", ".join(map(str, b["color_palette"]))) if bp: parts.append("Building: " + "; ".join(bp)) i = canon.get("interior") or {} if i: ip = [] for k in ("wall_finish", "floor_finish", "ceiling", "lighting_fixtures", "general_clutter_level"): v = i.get(k) if v: if isinstance(v, list): v = ", ".join(map(str, v)) ip.append(f"{k}={v}") if ip: parts.append("Interior: " + "; ".join(ip)) return ". ".join(parts).strip() VIEW_PROMPTS = [ ("N", "facing NORTH, looking toward the north wall of the room"), ("E", "facing EAST (90 degrees clockwise rotation from the previous N view), looking toward the east wall"), ("S", "facing SOUTH (180 degrees from the N view, opposite the entry), looking toward the south wall"), ("W", "facing WEST (270 degrees clockwise from the N view), looking toward the west wall"), ] def build_prompt(direction_desc: str, canon_text: str, is_first: bool) -> str: base = ( "Photorealistic cinematic interior eye-level still, 35mm film aesthetic. " "Same room as the architectural floor plan reference image. " f"Camera position: standing roughly at the room's center, {direction_desc}, " "approx 1.6m camera height (human eye-level), slight wide angle ~28mm.\n\n" "**THIS IS A DIFFERENT CAMERA ANGLE — show the wall and furniture that fall in " f"the {direction_desc.split(',')[0]} half of the room. The composition must clearly " "differ from any previous view photo (different walls, different furniture in frame).**\n\n" "Strict requirements:\n" "- Maintain EXACT wall layout, door positions, window positions, and furniture " "placement as shown in the floor plan reference. Do not invent walls or furniture.\n" "- Visual identity (wall finish, floor finish, ceiling, lighting fixtures, " "color palette, weathering) consistent with canon below " + ("AND with previous view photo references provided (but DIFFERENT camera angle)" if not is_first else "") + ".\n" "- No people, no characters, no figures, empty room, uninhabited.\n" "- Soft natural daylight from window mixed with dim domestic practicals " "(weak floor lamp, faint old television glow if visible). Realistic shadows.\n" "- Subtle depth of field, slight film grain, lived-in details " "(dust motes, scuff marks, subtle wear).\n" "- No overlaid text, no diagram lines, no architectural label letters, " "no captions, no watermarks.\n" "- 35-degree horizontal field of view fits one major wall and partial side walls." ) if canon_text: base += f"\n\nCanon: {canon_text}" return base def main() -> int: p = argparse.ArgumentParser() p.add_argument("--plan-png", required=True, help="interior floor plan PNG (reference)") p.add_argument("--out-dir", required=True) p.add_argument("--run-dir", default=None, help="(선택) v4 run dir — step1_spatial.json에서 canon 자동 추출") p.add_argument("--canon", default="", help="canon spec 영문 한 단락 (직접 지정)") p.add_argument("--model", default="gpt-image-2") p.add_argument("--size", default="1536x1024") p.add_argument("--quality", default="high") p.add_argument("--no-prev-chain", action="store_true", help="prev view photo를 ref에서 제외 (시점이 같아지는 문제 회피)") args = p.parse_args() plan_png = Path(args.plan_png).resolve() if not plan_png.exists(): logger.error("plan-png 없음: %s", plan_png); return 1 out_dir = Path(args.out_dir).resolve() out_dir.mkdir(parents=True, exist_ok=True) canon_text = args.canon if not canon_text and args.run_dir: sp_path = Path(args.run_dir) / "step1_spatial.json" if sp_path.exists(): spatial = json.loads(sp_path.read_text(encoding="utf-8")) canon_text = compact_canon_text(spatial.get("environment_canon") or {}) logger.info("canon from run-dir: %d chars", len(canon_text)) if not os.getenv("OPENAI_API_KEY"): logger.error("OPENAI_API_KEY not set"); return 1 client = OpenAI() prev_photo: Optional[Path] = None for tag, direction_desc in VIEW_PROMPTS: is_first = (prev_photo is None) or args.no_prev_chain prompt = build_prompt(direction_desc, canon_text, is_first) # refs: --no-prev-chain이면 plan만 (시점 다양화 우선), # 아니면 prev photo + plan (visual identity 우선) if args.no_prev_chain: refs = [plan_png] else: refs = ([prev_photo] if prev_photo else []) + [plan_png] logger.info("[view %s] refs=%d (prev=%s, plan=%s, no_chain=%s)", tag, len(refs), bool(prev_photo) and not args.no_prev_chain, plan_png.name, args.no_prev_chain) files = [open(p, "rb") for p in refs] try: resp = client.images.edit( model=args.model, image=files, prompt=prompt, size=args.size, quality=args.quality, n=1, ) finally: for f in files: f.close() out_p = out_dir / f"interior_view_{tag}.png" b64 = resp.data[0].b64_json if not b64: logger.error("[view %s] empty b64", tag); continue out_p.write_bytes(base64.b64decode(b64)) logger.info("[view %s] saved: %s (%d KB)", tag, out_p.name, out_p.stat().st_size // 1024) prev_photo = out_p logger.info("=== DONE === %s", out_dir) return 0 if __name__ == "__main__": sys.exit(main())