"""실험 — GPT-5.x가 씬 내용으로 4 view 순서/prompt를 직접 설계 → gpt-image-2 chain (도면 ref 없이). 흐름: Step A. run_dir + base_plan_id로 도면 + base_photo + 관련 shots 수집 Step B. GPT vision: 도면 PNG + scene → JSON {reasoning, atmosphere_canon, views[1..4]} views[i]: { order, label, description, t2i_prompt, uses_prev } Step C. gpt-image-2: - view 1: images.generate (텍스트 전용, reference 없음) - view 2,3,4: images.edit with [previous view image] as the ONLY reference (도면 ref 미사용) 가설: GPT가 prompt를 자기-완결형(self-contained)으로 만들어주면, 도면 없이도 chain만으로 방의 일관성을 유지할 수 있는지 검증. """ from __future__ import annotations import argparse import base64 import glob 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("gpt_planned_4view") SYSTEM_PROMPT = """You are an architectural visualization specialist designing a 4-view photorealistic chain of an interior room. Pipeline constraints (READ CAREFULLY): - Image 1 will be generated TEXT-ONLY by gpt-image-2 (NO reference image at all). - Images 2, 3, and 4 will be generated by gpt-image-2 image-edit with ONLY the immediately previous image as reference (NO floor plan reference, NO atmosphere image — just the previous photo). - Therefore each t2i_prompt must be self-contained: describe enough of the room's atmosphere, materials, and lighting that the model can produce a coherent photo without ever seeing the floor plan. Your job: 1. Decide the optimal ORDER of the 4 views to maximize cumulative consistency. Tips: - Image 1 should be the most distinctive/anchor-worthy direction (rich furniture, defining wall) so it pins the room identity. - Images 2-4 should rotate in a way that each view is spatially adjacent to the previous (shares a wall corner), so the model can reuse texture/lighting from the prev image naturally. - Avoid jumps that share no visible context (e.g., kitchen wall → opposite blank wall) — that breaks the chain. 2. For each view, output: - order: 1, 2, 3, or 4 - label: short snake_case name (e.g., "kitchen_wall", "bed_corner", "entry_door_wall") - description: 2-3 sentences in English describing the camera direction, what walls/doors/furniture are in frame - t2i_prompt: full single-string prompt for gpt-image-2 Prompt rules: - view 1 prompt: thoroughly describe the room atmosphere (wall finish, floor, ceiling, lighting, palette, lived-in details), plus the specific contents of this view (which doors/windows/furniture). 35mm cinematic still aesthetic, soft natural daylight + dim domestic practicals, realistic shadows, NO people, empty room. - view 2/3/4 prompt: must include * a brief atmospheric anchor ("same room as the previous photo — match its wall finish, floor, ceiling, lighting tone, color palette, and overall material aging") * a clear rotation cue ("camera rotated NN degrees clockwise from the previous view; now facing X") * the specific contents of THIS view (doors/windows/furniture), DIFFERENT from the previous image * 35mm cinematic still + same atmospheric details + NO people 3. Provide: - reasoning: ONE paragraph explaining why this order maximizes consistency - atmosphere_canon: ONE paragraph distilling the wall/floor/ceiling/lighting/palette canon, used internally and visible in every view's t2i_prompt STRICT RULES: - ENGLISH ONLY in all prompts and any rendered text. - COMMON NOUNS only — NEVER use proper names of characters, places, or productions from the scenes. - Reflect scene atmosphere (lived-in, worn) but NO narrative events: no people, no blood, no broken glass, no action. - Each view must show DIFFERENT content; no two views may look the same. - Mention the rotation step explicitly in views 2-4. Return strict JSON. No markdown fence, no commentary outside JSON. """ RESPONSE_SCHEMA = { "type": "object", "properties": { "reasoning": {"type": "string"}, "atmosphere_canon": {"type": "string"}, "views": { "type": "array", "minItems": 4, "maxItems": 4, "items": { "type": "object", "properties": { "order": {"type": "integer"}, "label": {"type": "string"}, "description": {"type": "string"}, "t2i_prompt": {"type": "string"}, "uses_prev": {"type": "boolean"}, }, "required": ["order", "label", "description", "t2i_prompt", "uses_prev"], "additionalProperties": False, }, }, }, "required": ["reasoning", "atmosphere_canon", "views"], "additionalProperties": False, } def collect_scene_context(run_dir: Path, base_plan_id: str) -> dict: ctx: dict = {"base_plan_id": base_plan_id, "base_plan": None, "base_photo": None, "shots": [], "spatial": None} sp2 = run_dir / "step2_plan_specs.json" if sp2.exists(): d = json.loads(sp2.read_text(encoding="utf-8")) for p in d.get("base_plans", []): if p.get("id") == base_plan_id: ctx["base_plan"] = p; break sp7 = run_dir / "step7_photo_specs.json" if sp7.exists(): d = json.loads(sp7.read_text(encoding="utf-8")) for ph in d.get("base_photos", []): if ph.get("source_plan_id") == base_plan_id: ctx["base_photo"] = ph; break for f in sorted(glob.glob(str(run_dir / "step3_shot_*.json"))): s = json.loads(Path(f).read_text(encoding="utf-8")) if s.get("base_plan_id") == base_plan_id: ctx["shots"].append(s) sp1 = run_dir / "step1_spatial.json" if sp1.exists(): ctx["spatial"] = json.loads(sp1.read_text(encoding="utf-8")) return ctx def build_user_prompt(ctx: dict) -> str: parts: list[str] = [] parts.append("== BASE PLAN SPEC ==") bp = ctx.get("base_plan") or {} parts.append(f"id: {bp.get('id')}") parts.append(f"label: {bp.get('label')}") parts.append(f"visual_domain: {bp.get('visual_domain')}") if bp.get("t2i_prompt"): parts.append(f"\nplan_drawing_prompt:\n{bp['t2i_prompt']}") if bp.get("legend"): parts.append(f"\nlegend: {json.dumps(bp['legend'], ensure_ascii=False)}") if bp.get("elements_meta"): parts.append(f"\nelements_meta: {json.dumps(bp['elements_meta'], ensure_ascii=False)}") parts.append("\n== BASE PHOTO SPEC (existing wall-by-wall annotation) ==") bph = ctx.get("base_photo") or {} if bph.get("camera_note"): parts.append(f"camera_note: {bph['camera_note']}") if bph.get("lighting"): parts.append(f"lighting: {bph['lighting']}") if bph.get("t2i_prompt"): parts.append(f"\nbase_photo_prompt:\n{bph['t2i_prompt']}") parts.append("\n== RELATED SHOTS ==") for s in ctx.get("shots", []): si = s.get("scene_index"); sx = s.get("shot_index") parts.append(f"\n--- Shot S{si:02d}_Shot{sx} ---") cam = s.get("camera") or {} parts.append(f"camera position: {cam.get('position', '(none)')}") parts.append(f"camera heading: {cam.get('heading', '(none)')}") parts.append(f"height/fov/lens: {cam.get('height')} / {cam.get('fov')} / {cam.get('lens_note')}") for a in s.get("additions", []): parts.append(f" + [{a.get('type')}] @ {a.get('position', '?')}: {a.get('note', '')}") parts.append("\n== INTERIOR CANON ==") sp = ctx.get("spatial") or {} interior = (sp.get("environment_canon") or {}).get("interior") or {} if interior: parts.append(json.dumps(interior, ensure_ascii=False, indent=2)) parts.append("\n== TASK ==") parts.append("Use the floor plan IMAGE (attached, for spatial reasoning ONLY) and ALL information above to plan the 4-view chain.") parts.append("Return JSON with: reasoning, atmosphere_canon, and 4 views (order 1..4 each with label/description/t2i_prompt/uses_prev).") return "\n".join(parts) def call_gpt_vision(client: OpenAI, model: str, plan_png: Path, system_prompt: str, user_prompt: str) -> dict: with open(plan_png, "rb") as f: b64 = base64.b64encode(f.read()).decode("ascii") resp = client.chat.completions.create( model=model, messages=[ {"role": "system", "content": system_prompt}, {"role": "user", "content": [ {"type": "text", "text": user_prompt}, {"type": "image_url", "image_url": {"url": f"data:image/png;base64,{b64}"}}, ]}, ], response_format={ "type": "json_schema", "json_schema": { "name": "FourViewChainPlan", "schema": RESPONSE_SCHEMA, "strict": True, }, }, ) txt = resp.choices[0].message.content or "" return json.loads(txt) def gpt_image_generate(client: OpenAI, model: str, prompt: str, size: str, quality: str, out_path: Path) -> Path: logger.info("[image generate] %s → %s (%s, %s)", model, out_path.name, size, quality) resp = client.images.generate( model=model, 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 gpt_image_edit_with_prev(client: OpenAI, model: str, prev_path: Path, prompt: str, size: str, quality: str, out_path: Path) -> Path: logger.info("[image edit] %s + prev[%s] → %s (%s, %s)", model, prev_path.name, out_path.name, size, quality) with open(prev_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("--run-dir", required=True) p.add_argument("--base-plan-id", required=True) p.add_argument("--out-dir", required=True) p.add_argument("--text-model", default=os.getenv("OPENAI_MODEL", "gpt-5.5")) p.add_argument("--image-model", default="gpt-image-2") p.add_argument("--size", default="1024x1024") p.add_argument("--quality", default="high") p.add_argument("--skip-text", action="store_true") p.add_argument("--text-only", action="store_true") args = p.parse_args() plan_png = Path(args.plan_png).resolve() run_dir = Path(args.run_dir).resolve() out_dir = Path(args.out_dir).resolve() out_dir.mkdir(parents=True, exist_ok=True) if not os.getenv("OPENAI_API_KEY"): logger.error("OPENAI_API_KEY not set"); return 1 client = OpenAI() prompts_p = out_dir / "prompts.json" # Step A + B if args.skip_text and prompts_p.exists(): plan = json.loads(prompts_p.read_text(encoding="utf-8")) logger.info("[skip text] reusing prompts.json") else: ctx = collect_scene_context(run_dir, args.base_plan_id) if not ctx["base_plan"]: logger.error("base_plan_id 매칭 실패: %s", args.base_plan_id); return 1 logger.info("collected: base_plan=ok, base_photo=%s, %d shots", bool(ctx["base_photo"]), len(ctx["shots"])) user_prompt = build_user_prompt(ctx) logger.info("user prompt: %d chars; calling %s", len(user_prompt), args.text_model) plan = call_gpt_vision(client, args.text_model, plan_png, SYSTEM_PROMPT, user_prompt) plan["_meta"] = { "base_plan_id": args.base_plan_id, "shot_count": len(ctx["shots"]), "text_model": args.text_model, "image_model": args.image_model, "size": args.size, "quality": args.quality, } prompts_p.write_text(json.dumps(plan, ensure_ascii=False, indent=2), encoding="utf-8") logger.info("saved prompts.json") # Print for review print("\n" + "=" * 60) print("=== REASONING ===") print(plan.get("reasoning", "(missing)")) print("\n=== ATMOSPHERE CANON ===") print(plan.get("atmosphere_canon", "(missing)")) views = sorted(plan.get("views", []), key=lambda v: v.get("order", 99)) for v in views: print(f"\n--- order={v.get('order')} label={v.get('label')} uses_prev={v.get('uses_prev')} ---") print(f"description: {v.get('description', '')}") print(f"t2i_prompt:\n{v.get('t2i_prompt', '')}") print() if args.text_only: logger.info("text-only mode → skip image generation"); return 0 # Step C: chain image generation prev_path: Path | None = None for v in views: i = v.get("order") label = v.get("label", f"view_{i}") prompt = v.get("t2i_prompt", "") if not prompt: logger.error("missing t2i_prompt for order=%s — skip", i); continue out_p = out_dir / f"{i:02d}_{label}.png" if prev_path is None: gpt_image_generate(client, args.image_model, prompt, args.size, args.quality, out_p) else: gpt_image_edit_with_prev(client, args.image_model, prev_path, prompt, args.size, args.quality, out_p) prev_path = out_p logger.info("=== DONE === %s", out_dir) return 0 if __name__ == "__main__": sys.exit(main())