"""S5_Shot3 chain bg 이중 묘사 가설 검증 테스트. 가설: chain bg PNG가 reference로 주입되면, 그 이미지에 이미 그려진 배경 요소 (TV, 창문, 싱크대, 식탁, 문)을 t2i_prompt가 또 묘사할 때 모델이 그 자리에 이중 합성을 하거나 배경을 변형시킨다. 검증 방법: 같은 reference 이미지 + 같은 모델로 prompt만 두 버전 호출: A. 기존 prompt (재묘사 포함, production 그대로) B. 기존 prompt + [배경 참조 안내] 섹션 추가 ("TV/창문/싱크대 등은 chain bg에 이미 있음. 그대로 유지하고 새로 그리지 말 것") production 코드 (`gemini_i2i_editor._gemini_generate_content`)와 동일한 방식 으로 호출. 모델은 nano-banana-2 (gemini-3.1-flash-image-preview). reference 순서는 production resolve_refs_for_prompt와 동일: 1. chain bg ref (background_chain_node primary path) 2. character C04 primary ref (composite 없으므로 face only) 3. P06 prop ref (t2i_prompt에 P06 명시됨) 미포함: - O06 outfit (composite 없어 production이 차단) - P03 prop (t2i_prompt에 P03 미명시) - prev_shot_ref (chain bg 있어 fallback 안 됨) - shot_dependency (S5_Shot3 dep_id=None) 사용법: cd backend .venv/bin/python scripts/experiment_chain_bg_double_render_test.py 결과: scripts/experiment_results/s5shot3_A_original.png scripts/experiment_results/s5shot3_B_with_bg_note.png scripts/experiment_results/s5shot3_metadata.json """ 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 # .env 로드 try: from dotenv import load_dotenv load_dotenv(Path(__file__).parent.parent / ".env") except ImportError: pass PID = "c00bbe19-a9b5-463f-acfc-806f2e820258" EID = "fe165e3a-19c2-4a0f-9acb-e0c9bab0ee5a" PROJECT_ROOT = Path("/Users/manta/Documents/Projects/TheRoad-I1") # Production reference 경로 (S5_Shot3 분석 결과) CHAIN_BG = ( PROJECT_ROOT / "projects" / PID / "images" / EID / "background_chain" / "L05" / "interior_living_kitchen_day_normal.png" ) 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" ) # 기존 production prompt (DB에서 가져온 그대로) 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." ) # 같은 prompt + 배경 참조 안내 섹션 추가 BG_NOTE = ( "\n\n[Background reference notice — read FIRST]\n" "The first reference image (background chain ref) already contains the full " "interior layout: the television (left side), the small window, the worn " "sink, the kitchen counter, the low dining table, the front door, and the " "wall finish. These background elements are FIXED and must be reused as-is " "from the reference image. DO NOT redraw them, do not change their " "position, size, color, or style. The prompt mentions some of these " "elements only to indicate where the character looks or what light hits " "her face — they are NOT instructions to redraw the elements themselves. " "Add only: the character (C04O06), her pose, her gaze direction, the " "steam, the steaming pot in the foreground, and P06 (a herb) beside the " "sink. Keep everything else identical to the background reference." ) PROMPT_B_WITH_NOTE = PROMPT_A_ORIGINAL + BG_NOTE # Production reference 라벨 (resolve_refs_for_prompt 순서 그대로) INPUT_IMAGES = [ ( "background chain ref (interior_living_kitchen_day_normal for L05) — " "match wall/floor/ceiling/lighting", CHAIN_BG.read_bytes(), ), ("character C04 identity", C04_FACE_REF.read_bytes()), ("object P06", P06_REF.read_bytes()), ] GEMINI_API_URL_TEMPLATE = ( "https://generativelanguage.googleapis.com/v1beta/models/{model}:" "generateContent?key={api_key}" ) MODEL = "gemini-3.1-flash-image-preview" ASPECT_RATIO = "16:9" TIMEOUT = 180 MAX_RETRIES = 2 def gemini_generate(api_key: str, prompt: str, input_images: list) -> bytes: """production gemini_i2i_editor._gemini_generate_content와 동일한 호출.""" 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=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_error = 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_error = 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_error from exc except (urllib.error.URLError, socket.timeout) as exc: last_error = exc if attempt <= MAX_RETRIES: time.sleep(2 * attempt) continue raise RuntimeError(f"URL error after retries: {exc}") from exc else: raise RuntimeError(f"Failed: {last_error}") pf = payload.get("promptFeedback", {}) if pf.get("blockReason"): raise RuntimeError(f"Moderation blocked: {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)[:500]}") def main(): api_key = os.environ.get("GEMINI_API_KEY", "") if not api_key: print("ERROR: GEMINI_API_KEY env not set", file=sys.stderr) sys.exit(1) out_dir = Path(__file__).parent / "experiment_results" out_dir.mkdir(exist_ok=True) print("=" * 60) print("S5_Shot3 chain bg 이중 묘사 가설 검증") print("=" * 60) print(f"Model: {MODEL}") print(f"References ({len(INPUT_IMAGES)}):") for i, (label, img) in enumerate(INPUT_IMAGES, 1): print(f" [{i}] {label[:60]}... ({len(img):,} bytes)") print() # A: 기존 prompt print("--- A: 기존 prompt (재묘사 포함) ---") print(f"prompt length: {len(PROMPT_A_ORIGINAL)} chars") a_path = out_dir / "s5shot3_A_original.png" try: a_bytes = gemini_generate(api_key, PROMPT_A_ORIGINAL, INPUT_IMAGES) a_path.write_bytes(a_bytes) print(f"OK: {a_path} ({len(a_bytes):,} bytes)") except Exception as exc: print(f"FAIL: {exc}") a_path = None print() # B: 배경 안내 추가 print("--- B: 기존 + [배경 참조 안내] 섹션 ---") print(f"prompt length: {len(PROMPT_B_WITH_NOTE)} chars") b_path = out_dir / "s5shot3_B_with_bg_note.png" try: b_bytes = gemini_generate(api_key, PROMPT_B_WITH_NOTE, INPUT_IMAGES) b_path.write_bytes(b_bytes) print(f"OK: {b_path} ({len(b_bytes):,} bytes)") except Exception as exc: print(f"FAIL: {exc}") b_path = None print() # metadata meta = { "test": "S5_Shot3 chain bg double-render verification", "project_id": PID, "episode_id": EID, "model": MODEL, "aspect_ratio": ASPECT_RATIO, "references": [ {"label": "chain bg", "path": str(CHAIN_BG)}, {"label": "C04 face", "path": str(C04_FACE_REF)}, {"label": "P06 prop", "path": str(P06_REF)}, ], "variants": { "A_original": { "prompt": PROMPT_A_ORIGINAL, "result_path": str(a_path) if a_path else None, }, "B_with_bg_note": { "prompt": PROMPT_B_WITH_NOTE, "result_path": str(b_path) if b_path else None, "added_section": BG_NOTE, }, }, } meta_path = out_dir / "s5shot3_metadata.json" meta_path.write_text( json.dumps(meta, ensure_ascii=False, indent=2), encoding="utf-8" ) print(f"metadata: {meta_path}") print() print("비교 항목:") print(" 1. TV 이중 묘사: A vs B에서 TV 영역 변형/이중 가구 발생 여부") print(" 2. 창문/싱크대 위치: chain bg 그대로 유지 여부") print(" 3. 인물 자세 + 시선 + 김 + 냄비: 두 버전 모두 정상 합성") print() print("원본 chain bg + production scene 이미지:") print(f" chain bg: {CHAIN_BG}") print( f" production: {PROJECT_ROOT}/projects/{PID}/images/{EID}/scene/" "cb2a0a73-96a7-43d2-b5f4-796f3b66ce6e.png" ) if __name__ == "__main__": main()