"""실험 — interior 도면 reference로 (1) 파노라마 1장 + (2) 4분할(N/E/S/W) 1장 생성. GPT-image-2 단일. 도면 PNG만 reference. 사용: python scripts/experiment_panorama_and_quad.py --plan-png --out-dir [--canon "..."] """ from __future__ import annotations import argparse import base64 import json import logging import os import sys from pathlib import Path 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("pano_quad") 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() def panorama_prompt(canon_text: str) -> str: return ( "Photorealistic 360-degree EQUIRECTANGULAR INTERIOR PANORAMA of the room shown in the " "architectural floor plan reference image. The output is a single horizontal strip image " "where all 4 walls of the room are visible side by side as if the camera spun a full 360 " "degrees from the room's center.\n\n" "Strict requirements:\n" "- Equirectangular panoramic projection (typical for 360 viewers): wide aspect ratio with " "natural perspective foreshortening near top/bottom edges, horizontally stitched seamless.\n" "- Camera position: standing at the geometric center of the room, eye-level approx 1.6m height.\n" "- Reading order LEFT to RIGHT corresponds to camera rotating clockwise: NORTH wall first " "(left edge) → EAST wall (left-center) → SOUTH wall (center-right) → WEST wall (right edge), " "wrapping back to NORTH at the seam.\n" "- Maintain EXACT wall layout, door positions, window positions, and furniture placement " "from the floor plan reference. Do not invent walls or furniture.\n" "- Photorealistic, 35mm cinematic still aesthetic. Soft natural daylight from windows mixed " "with dim domestic practicals (weak floor lamp, faint television glow). Realistic shadows.\n" "- Lived-in details: dust motes, scuff marks, subtle wear.\n" "- No people, no characters, empty room.\n" "- No overlaid text, no diagram lines, no architectural label letters, no captions.\n\n" f"Canon: {canon_text or 'aged residential interior, modest domestic clutter'}" ) def quad_prompt(canon_text: str) -> str: return ( "Photorealistic 4-QUADRANT COMPOSITE image showing 4 different eye-level camera views of " "the SAME interior room (matching the architectural floor plan reference), arranged in a " "2x2 grid with thin black dividers between quadrants.\n\n" "Layout (2x2 grid):\n" "- TOP-LEFT quadrant: camera facing NORTH wall, label 'N' in top-left corner of this quadrant.\n" "- TOP-RIGHT quadrant: camera facing EAST wall (90 degrees clockwise from N), label 'E'.\n" "- BOTTOM-LEFT quadrant: camera facing WEST wall (270 degrees from N), label 'W'.\n" "- BOTTOM-RIGHT quadrant: camera facing SOUTH wall (180 degrees from N), label 'S'.\n\n" "Each quadrant must:\n" "- Be a distinct camera angle showing different walls/furniture (NOT identical compositions).\n" "- Camera at room center, eye-level approx 1.6m height, slight wide angle ~28mm.\n" "- Maintain EXACT wall layout, door, window, furniture placement from floor plan reference.\n" "- Visual identity (wall finish, floor finish, ceiling, lighting, color palette) " "consistent across all 4 quadrants — same room, same lighting, just different angles.\n" "- Photorealistic 35mm cinematic still, soft natural daylight + dim practicals.\n" "- No people, empty room.\n" "- No overlaid text inside quadrants, no diagram lines (other than the grid divider).\n" "- Quadrant corner labels (N/E/W/S) in ENGLISH ONLY, small clean sans-serif.\n\n" f"Canon: {canon_text or 'aged residential interior, modest domestic clutter'}" ) def edit_with_plan(client: OpenAI, model: str, plan_path: Path, prompt: str, size: str, quality: str, out_path: Path) -> Path: logger.info("[image edit] %s + plan ref → %s (%s, %s)", model, out_path.name, size, quality) with open(plan_path, "rb") as f: resp = client.images.edit( model=model, image=[f], prompt=prompt, size=size, quality=quality, n=1, ) b64 = resp.data[0].b64_json if not b64: raise RuntimeError(f"empty b64 for {out_path.name}") out_path.write_bytes(base64.b64decode(b64)) logger.info(" saved %d KB", out_path.stat().st_size // 1024) return out_path def main() -> int: p = argparse.ArgumentParser() p.add_argument("--plan-png", required=True) p.add_argument("--out-dir", required=True) p.add_argument("--run-dir", default=None, help="(선택) v4 run dir에서 canon 자동 추출") p.add_argument("--canon", default="") p.add_argument("--model", default="gpt-image-2") p.add_argument("--panorama-size", default="1792x1024", help="파노라마 크기 (가로 긴 비율). 1792x1024 / 2048x1024 등") p.add_argument("--quad-size", default="1024x1024", help="4분할 크기 (정사각)") p.add_argument("--quality", default="high") args = p.parse_args() plan_png = Path(args.plan_png).resolve() if not plan_png.exists(): logger.error("plan-png 없음"); 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(args.run_dir) / "step1_spatial.json" if sp.exists(): spatial = json.loads(sp.read_text(encoding="utf-8")) canon_text = compact_canon_text(spatial.get("environment_canon") or {}) logger.info("canon: %d chars", len(canon_text)) if not os.getenv("OPENAI_API_KEY"): logger.error("OPENAI_API_KEY not set"); return 1 client = OpenAI() # 1) Panorama pano_p = out_dir / "interior_panorama.png" try: edit_with_plan(client, args.model, plan_png, panorama_prompt(canon_text), args.panorama_size, args.quality, pano_p) except Exception as e: logger.error("panorama failed: %s", e) # 2) Quad quad_p = out_dir / "interior_quad_NESW.png" try: edit_with_plan(client, args.model, plan_png, quad_prompt(canon_text), args.quad_size, args.quality, quad_p) except Exception as e: logger.error("quad failed: %s", e) # prompts 저장 (out_dir / "prompts.json").write_text( json.dumps({ "panorama": panorama_prompt(canon_text), "quad": quad_prompt(canon_text), "canon": canon_text, }, ensure_ascii=False, indent=2), encoding="utf-8" ) logger.info("=== DONE === %s", out_dir) return 0 if __name__ == "__main__": sys.exit(main())