"""S5_Shot3 chain bg — floor plan v2 (TV ↔ 소파 마주봄, variant E). 이전 도면(00_floorplan.png)의 결함: - "Left wall: TV ... a fabric sofa parallel to the wall facing the TV" → gpt-image-2가 'TV + 소파 둘 다 left wall에 붙임'으로 해석 - 결과: 도면 자체가 이상 + chain bg에서 TV가 싱크대 옆으로 이동 v2 fix — 명확하게 분리: - 좌측 벽: TV만 (벽에 붙음, 화면이 오른쪽 향함) - 우측 벽: 소파만 (벽에 붙음, 좌측 향함) - 즉 TV ↔ 소파는 방 가운데를 사이에 두고 마주봄 - 부엌(싱크대+창)은 뒤쪽 벽 (back wall) - 식탁은 뒤쪽 벽 가까이 또는 가운데 Step E1: v2 도면 생성 Step E2: v2 도면 기반 chain bg 재생성 Step E3: 새 chain bg + v2 도면 + 인물/소품 + 기존 prompt(A) → variant E """ from __future__ import annotations import base64 import json import os import socket import sys import time import urllib.error import urllib.request from pathlib import Path try: from dotenv import load_dotenv load_dotenv(Path(__file__).parent.parent / ".env") except ImportError: pass from openai import OpenAI PID = "c00bbe19-a9b5-463f-acfc-806f2e820258" EID = "fe165e3a-19c2-4a0f-9acb-e0c9bab0ee5a" PROJECT_ROOT = Path("/Users/manta/Documents/Projects/TheRoad-I1") C04_FACE_REF = ( PROJECT_ROOT / "projects" / PID / "images" / EID / "reference" / "008b5d4c-5a02-4233-853e-ffc99225cfe1.png" ) P06_REF = ( PROJECT_ROOT / "projects" / PID / "images" / EID / "reference" / "48735792-33c0-415b-afdf-46146fe19ed4.png" ) OUT_DIR = Path(__file__).parent / "experiment_results" # Step E1: v2 floor plan prompt — TV/소파 명확 분리 FLOORPLAN_V2_PROMPT = ( "Top-down architectural floor plan diagram, hand-drawn black ink lines on " "off-white paper, single small Korean rooftop apartment (옥탑방) interior " "viewed directly from above. Single rectangular room, roughly 5m wide x " "5m deep. NO interior walls dividing the room. " "Label and place each item EXACTLY where described — do not combine items " "on the same wall: " "(1) LEFT WALL: ONE television on a low stand, drawn as a small rectangle " "against the left wall, screen facing rightward into the room. Label: " "'TV'. Nothing else on the left wall. " "(2) RIGHT WALL: ONE fabric sofa drawn as a long rectangle against the " "right wall, with seat side facing leftward into the room. Label: 'SOFA'. " "The TV and the sofa are on OPPOSITE walls and face each other across the " "empty middle of the room. " "(3) BACK WALL (top edge of plan): kitchen counter running along the back " "wall. In the middle of this counter, draw a sink (square with faucet " "symbol). Label: 'SINK'. To the right of the sink along the same back " "wall, draw a small window with curtain symbol. Label: 'WINDOW'. At the " "far back-right corner of the room, draw a small interior doorway opening " "to a tiny adjacent space. Label: 'BEDROOM DOOR'. " "(4) FRONT WALL (bottom edge of plan, where the camera enters): in the " "middle of the front wall, draw the front entry door swinging outward " "(arc symbol showing swing). Label: 'FRONT DOOR'. " "(5) BETWEEN SOFA AND SINK (in the open floor area, NOT against any wall): " "draw a simple wooden dining table with two chairs. Label: 'TABLE'. " "Floor surface: wooden floorboards indicated by light parallel line " "hatching across the entire room. " "Style: clean schematic architectural diagram, plain black ink linework " "on warm off-white background, English labels in small caps, no " "perspective, no shading, no people, no furniture grouped together — " "each item must be clearly separated and clearly placed against the wall " "described." ) # Step E2: v2 도면 기반 chain bg 재생성 prompt CHAIN_BG_V2_PROMPT = ( "Photorealistic cinematic interior background. Convert the architectural " "floor plan reference (top-down diagram) into a 35mm film-style photoreal " "view from inside the room, eye-level wide angle, single static frame, " "no people. Camera is positioned NEAR THE FRONT-LEFT CORNER looking " "diagonally toward the back-right corner so that BOTH the left wall (with " "the TV) and the back wall (with the sink and window) are clearly visible " "in the frame, and the right side shows the sofa across the room and the " "dining table. " "Match the floor plan layout EXACTLY: " "ONE television on a low wooden stand against the left wall, screen " "facing into the room (toward the sofa across the empty floor). " "ONE fabric sofa against the right wall, facing leftward toward the TV. " "Kitchen counter with stainless sink in the middle of the back wall, a " "small window with curtain to the right of the sink along the same back " "wall. A narrow plain interior doorway (closed) at the back-right corner " "leading to a tiny adjoining bedroom space (do NOT show its inside). " "A simple wooden dining table with two chairs in the middle of the floor " "between the sofa and the kitchen counter. " "Front entry door is offscreen behind the camera, ignore it. " "This is ONE single rooftop room (옥탑방), NOT a multi-room apartment. " "DRAW ONLY ONE TV (do NOT add a second TV anywhere). " "Material/lighting: dirty plaster walls in muted yellow-gray, worn wooden " "floorboards across the whole room, low ceiling with a single bare bulb " "fixture, pale daylight leaking through the small back window, soft warm " "tungsten ambient light filling the rest, dust in the air, damp domestic " "shadows, uneasy stillness, 35mm film grain, muted blue-gray and washed " "yellow palette." ) # Step E3: 기존 prompt (A 그대로 — 안내 X) PROMPT_A_ORIGINAL = ( "Photorealistic cinematic still. " "[L05: A cramped modern Korean 옥탑방 (rooftop room) interior, dusty small " "window, worn sink, low table, pale daylight, damp muted domestic shadows.] " "C04O06 in a plain apron occupies the right-center of the frame, at " "three-quarter angle from the left, one hand gripping the apron hem, eyes " "fixed on the television at the far left. The lens emphasizes her serious " "expression, tightened jaw, and the tense hand at the apron hem through a " "thin veil of white kitchen steam in the left foreground. A blurred " "television edge, screen angled away, sits at the far left, its cold " "blue-gray glow brushing her cheek; pale daylight from the closed dusty " "window adds washed yellow highlights. A steaming pot with a slightly " "lifted lid sits in the foreground, P06 rests rinsed beside the sink, and " "a thin stream from the faucet falls into the basin in the background. " "Muted blue-gray and washed yellow palette, damp domestic surfaces, uneasy " "stillness." ) GEMINI_API_URL_TEMPLATE = ( "https://generativelanguage.googleapis.com/v1beta/models/{model}:" "generateContent?key={api_key}" ) GEMINI_MODEL = "gemini-3.1-flash-image-preview" ASPECT_RATIO = "16:9" TIMEOUT = 240 MAX_RETRIES = 2 def gemini_generate(api_key: str, prompt: str, input_images: list) -> bytes: parts: list = [{"text": prompt}] for label, img_bytes in input_images: if label: parts.append({"text": label}) parts.append({ "inline_data": { "mime_type": "image/png", "data": base64.b64encode(img_bytes).decode("ascii"), } }) body = { "contents": [{"parts": parts}], "generationConfig": { "responseModalities": ["TEXT", "IMAGE"], "imageConfig": {"aspectRatio": ASPECT_RATIO, "imageSize": "2K"}, }, } url = GEMINI_API_URL_TEMPLATE.format(model=GEMINI_MODEL, api_key=api_key) req = urllib.request.Request( url, data=json.dumps(body).encode("utf-8"), headers={"Content-Type": "application/json"}, method="POST", ) last = None for attempt in range(1, MAX_RETRIES + 2): try: with urllib.request.urlopen(req, timeout=TIMEOUT) as resp: payload = json.loads(resp.read().decode("utf-8")) break except urllib.error.HTTPError as exc: text = exc.read().decode("utf-8", errors="replace") last = RuntimeError(f"HTTP {exc.code}: {text}") if exc.code in {429, 500, 502, 503, 504} and attempt <= MAX_RETRIES: time.sleep(2 * attempt); continue raise last from exc except (urllib.error.URLError, socket.timeout) as exc: last = exc if attempt <= MAX_RETRIES: time.sleep(2 * attempt); continue raise RuntimeError(f"URL error: {exc}") from exc else: raise RuntimeError(f"Failed: {last}") pf = payload.get("promptFeedback", {}) if pf.get("blockReason"): raise RuntimeError(f"Moderation: {pf['blockReason']}") cands = payload.get("candidates", []) if cands and cands[0].get("finishReason") == "SAFETY": raise RuntimeError("Safety filter") for cand in cands: for part in cand.get("content", {}).get("parts", []): inline = part.get("inlineData") or part.get("inline_data") if isinstance(inline, dict) and inline.get("data"): return base64.b64decode(inline["data"]) raise RuntimeError(f"No image: {json.dumps(payload)[:300]}") def main(): openai_key = os.environ.get("OPENAI_API_KEY", "") gemini_key = os.environ.get("GEMINI_API_KEY", "") if not openai_key or not gemini_key: print("ERROR: OPENAI_API_KEY + GEMINI_API_KEY required", file=sys.stderr) sys.exit(1) OUT_DIR.mkdir(exist_ok=True) client = OpenAI(api_key=openai_key) # ── Step E1: v2 도면 생성 ── print("=" * 60) print("Step E1: floor plan v2 (gpt-image-2 text-to-image)") print("=" * 60) print(f"prompt length: {len(FLOORPLAN_V2_PROMPT)} chars") floorplan_v2_path = OUT_DIR / "20_floorplan_v2.png" if floorplan_v2_path.exists(): print(f"이미 존재 — 재사용: {floorplan_v2_path}") else: resp = client.images.generate( model="gpt-image-2", prompt=FLOORPLAN_V2_PROMPT, size="1024x1024", quality="high", n=1, ) b64 = resp.data[0].b64_json floorplan_v2_path.write_bytes(base64.b64decode(b64)) print(f"OK: {floorplan_v2_path} ({floorplan_v2_path.stat().st_size:,} bytes)") print() # ── Step E2: v2 도면 기반 chain bg 재생성 ── print("=" * 60) print("Step E2: chain bg v2 (gpt-image-2 + v2 도면 ref)") print("=" * 60) print(f"prompt length: {len(CHAIN_BG_V2_PROMPT)} chars") new_chain_bg_v2_path = OUT_DIR / "21_new_chain_bg_v2.png" if new_chain_bg_v2_path.exists(): print(f"이미 존재 — 재사용: {new_chain_bg_v2_path}") else: with open(floorplan_v2_path, "rb") as f: resp = client.images.edit( model="gpt-image-2", image=f, prompt=CHAIN_BG_V2_PROMPT, size="1536x1024", quality="high", n=1, ) b64 = resp.data[0].b64_json new_chain_bg_v2_path.write_bytes(base64.b64decode(b64)) print(f"OK: {new_chain_bg_v2_path} ({new_chain_bg_v2_path.stat().st_size:,} bytes)") print() # ── Step E3: variant E — 새 chain bg + v2 도면 + 인물/소품 + 기존 prompt(A) ── print("=" * 60) print("Step E3: variant E — chain bg v2 + 도면 v2 + 인물/소품 + prompt A") print("=" * 60) input_images = [ ( "spatial layout reference (floor plan, top-down) — " "TV on left wall, sofa on right wall, they FACE EACH OTHER across " "the empty middle of the room", floorplan_v2_path.read_bytes(), ), ( "background reference (rebuilt photoreal interior matching the " "floor plan) — TV against left wall, sofa against right wall, " "kitchen and window on back wall; reuse all furniture positions " "as drawn here", new_chain_bg_v2_path.read_bytes(), ), ("character C04 identity", C04_FACE_REF.read_bytes()), ("object P06", P06_REF.read_bytes()), ] print(f"References (4):") for i, (label, _) in enumerate(input_images, 1): print(f" [{i}] {label[:80]}...") print(f"prompt length: {len(PROMPT_A_ORIGINAL)} chars (A 그대로)") print() e_path = OUT_DIR / "s5shot3_E_floorplan_v2.png" try: e_bytes = gemini_generate(gemini_key, PROMPT_A_ORIGINAL, input_images) e_path.write_bytes(e_bytes) print(f"OK: {e_path} ({len(e_bytes):,} bytes)") except Exception as exc: print(f"FAIL: {exc}") e_path = None print() meta = { "test": "S5_Shot3 — floor plan v2 (TV/sofa opposite walls), variant E", "step1_floor_plan_v2": { "model": "gpt-image-2", "prompt": FLOORPLAN_V2_PROMPT, "result_path": str(floorplan_v2_path), }, "step2_chain_bg_v2": { "model": "gpt-image-2", "input_ref": str(floorplan_v2_path), "prompt": CHAIN_BG_V2_PROMPT, "result_path": str(new_chain_bg_v2_path), }, "step3_variant_E": { "model": GEMINI_MODEL, "references": [ "20_floorplan_v2.png", "21_new_chain_bg_v2.png", "C04 face", "P06 prop", ], "prompt": PROMPT_A_ORIGINAL, "result_path": str(e_path) if e_path else None, }, } (OUT_DIR / "s5shot3_floorplan_v2_metadata.json").write_text( json.dumps(meta, ensure_ascii=False, indent=2), encoding="utf-8" ) print(f"metadata: {OUT_DIR / 's5shot3_floorplan_v2_metadata.json'}") if __name__ == "__main__": main()