#!/usr/bin/env python3 """s34 — 연속성 콘티 시트 실험 (실험 전용·커밋 금지). v22(2026-07-12): 영어 계약 원복(한국어 전문은 타 그룹 품질 저하로 반려) + 비유·은유·과장 캐릭터 표현의 실촬영 현실화 조항(REALIZE) 추가. v21(반려): 프롬프트 전문 한국어 원어 — 숲 인영은 해결했으나 그 외 악화. v20: 실사영화 프리프로덕션 스토리보드 스타일(만화체 금지, 영어 계약). 사용자 정정(v2): 콘티 = 손으로 그린 간략 러프 스케치 (실사 금지). 목적 = 구도·장면 연속성(인물 이동 방향/카메라 움직임/카메라 구도)만. 글자 절대 금지(샷 번호만 허용). 분할은 2/4/6 균등 그리드만. 패널 수 = 샷 수 그대로 — 홀수면 남는 패널을 순백으로 비움. 균등 분할이라 코드가 나중에 패널을 잘라 분해할 수 있어야 함. 사용: backend/.venv/bin/python s34_conti_sheet.py --only [--engines i2,nb2] [--groups forest,...] 산출: out/conti/conti2_*.png + plans/s34_conti_v2.json + conti_sheet.html """ import argparse import html as _html import json import sys from pathlib import Path sys.path.insert(0, str(Path(__file__).parent)) import forest_lib as F # noqa: E402 OUTD = F.OUT / "conti" PAGE = F.EXP / "conti_sheet.html" # v22: 영어 계약 원복 — 플랜 v2로 복귀(영어 동선 28건 보존분 재사용), v3=한국어판 보존 PLAN = "s34_conti_v2" PLAN_V1 = "s34_conti_v1" DATA = json.load(open(Path(__file__).parent / "pipeline_doc" / "shot_loc_data.json")) SHOT = {(s["scene"], s["shot"]): s for s in DATA["shots"]} ENGINES = ["i2", "nb2"] # (키, 표시명, 장소 서술 EN, [(scene, shot)...]) GROUPS = [ ("forest", "숲속", "a dense Korean mountain forest with thick undergrowth", [(1, 1), (1, 4), (3, 3)]), ("busstop", "해안가 버스정류장", "a small seaside bus stop with a bench facing the sea at night on a quiet Korean coastal road", [(15, 2), (15, 6)]), ("rooftop_ext", "옥탑 단지 외부", "the exterior of an old Korean multi-family villa: narrow alley and small yard at ground level, an added exterior steel staircase climbing the wall, and a rooftop with a small rooftop room (steel front door, one lit window), a water tank, a clothesline and a low parapet", [(10, 6), (10, 7), (11, 4), (11, 5), (13, 4), (13, 6)]), ("rooftop_int", "옥탑방 내부", "the interior of a small Korean rooftop home: living room with a small window, kitchenette and dining table; a bedroom with a bed against one wall and a curtained window; a tiny tiled bathroom with a mirror", [(12, 11), (12, 21), (14, 5), (18, 11), (25, 1), (25, 5)]), ("mart", "대형마트", "a mid-size Korean supermarket: shelved sales floor inside; the entrance area outside with a bicycle rack and one wall-mounted CCTV camera above; a small back security office with CCTV monitors", [(8, 3), (8, 5), (19, 5), (19, 6), (20, 1), (20, 3)]), ("police", "경찰서", "a modest Korean provincial police station: a detectives' office with desks, monitors and a doorway; and the street right in front of the building", [(16, 3), (21, 7), (21, 10), (21, 12), (22, 1)]), ] # 그리드: 샷 수 → (패널 슬롯 수, 등분법 명시 EN, i2 size, nb2 aspect) def _grid(n_shots): if n_shots <= 2: return (2, "2 panels made by ONE vertical cut at exactly 50% of the image" " width — each panel is exactly half the width and the FULL" " height (1 row x 2 columns)", "1536x1024", "16:9") if n_shots <= 4: return (4, "4 panels made by ONE vertical cut at exactly 50% of the width" " and ONE horizontal cut at exactly 50% of the height — a 2x2" " grid, each panel exactly one quarter of the image", "1024x1024", "1:1") return (6, "6 panels made by TWO vertical cuts at exactly 1/3 and 2/3 of the" " width and ONE horizontal cut at exactly 50% of the height — a" " 3-column x 2-row grid, all six panels exactly equal", "1536x1024", "16:9") HEAD = "\n".join([ "Draw ONE storyboard (conti) sheet: {layout}.", "STYLE: hand-drawn PRE-PRODUCTION STORYBOARD FOR A LIVE-ACTION FILM", "SHOOT — the working drawings a professional film storyboard artist", "makes for the director and cinematographer of a real movie:", "REALISTIC adult human proportions and faces (ordinary Korean", "adults, film-actor realism), true cinematic lens framing, and", "naturalistic light-and-shadow blocking rendered in monochrome", "graphite pencil tone on white paper.", "ABSOLUTELY NOT manga, NOT anime, NOT cartoon, NOT webtoon, NOT", "comic-book style: no stylized or enlarged eyes, no cute or", "youthful idealized faces, no clean ink outlines — every panel", "must read like a frame of a live-action film sketched in pencil.", "No color anywhere.", ]) PURPOSE = "\n".join([ "PURPOSE: each panel captures only the CAMERA COMPOSITION of one", "shot — framing (close-up / medium / wide), where the subject sits in", "the frame, and continuity between panels. Keep every panel", "extremely simple: only what matters for composition.", "MOVEMENT IS SHOWN BY THE FIGURE ITSELF, NEVER BY MARKINGS: convey", "direction of travel purely through the body — stride mid-motion,", "body lean into the path, back view vs front view, and the stated", "size/depth cues. Do NOT draw arrows, motion lines, path lines,", "speed lines or any diagram markings of any kind.", ]) REALIZE = "\n".join([ "REALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot", "text may describe characters metaphorically, figuratively or with", "exaggeration. Draw what a real movie camera would actually record", "on a physical set — exaggerated or figurative impressions become", "realistic staging choices (lighting, distance, angle, wardrobe),", "not literal fantasy drawings.", "EVERY CHARACTER IS A HUMAN BEING: unless the story explicitly", "features non-human or virtual beings (as in science fiction or", "fantasy), every character — however indirectly, vaguely or", "figuratively the text describes them — IS a real human. When the", "text gives no direct visual description of a person, IMAGINE one", "and still draw them as a concrete, fully-formed human being:", "always render the human form to the maximum extent — build,", "posture, face, hands, clothing — never reduce a person to a shape,", "blob, solid silhouette or abstract mass.", ]) CONSISTENCY = "\n".join([ "CONTINUITY: all drawn panels are the SAME single location — fixed", "features (doors, stairs, windows, poles, desks) keep the same", "position and left/right relationships across panels so the panels", "cut together as one continuous scene.", ]) GEOMETRY = "\n".join([ "GEOMETRY — ANALYSE BEFORE DRAWING: from each panel's Korean shot", "text, first work out the spatial facts and draw EXACTLY those, never", "re-stage them: (a) each figure's DISTANCE from camera (tight insert /", "close-up / medium / full / wide — a shot about a hand or an object", "is a tight insert, not a two-shot); (b) each figure's FACING and", "MOVEMENT DIRECTION relative to camera — '뒷모습' means seen from", "BEHIND, moving AWAY from camera into the depth of the frame, not", "sideways; a profile means lateral movement; (c) relative positions", "and distances BETWEEN figures. Keep every established screen", "direction consistent across panels (180-degree rule): a person who", "leaves in one panel keeps the same on-screen direction wherever the", "movement appears again.", "SIZE & DEPTH ARE CONTRACTUAL: each panel states every figure's", "apparent size in the frame and the explicit near/far depth order.", "Draw exactly those proportions — a farther figure MUST be drawn", "clearly smaller than a nearer one; never equalize figure sizes,", "especially when several figures share a panel or their paths", "relate to each other.", "DEPARTING FIGURE — a figure running/walking AWAY must be", "unmistakably departing, never approaching: seen exactly from behind", "(back of the head only, no face, no chin, no cheek line), body", "leaning into the path, mid-stride with one shoe sole toward the", "camera. THE STATED SIZE/DEPTH WINS: draw the departing figure at", "exactly the apparent size and depth the panel's FIGURES line states", "(if stated FARTHER/small, it stays small in the distance — never", "enlarge it into the foreground for drama).", ]) SEAL = "\n".join([ "ABSOLUTELY NO TEXT anywhere: no words, letters, captions, labels,", "sound effects or watermarks — the ONLY allowed marking is a small", "shot number digit ({numbers}) in a corner INSIDE each drawn panel.", "THE PANEL GRID FILLS THE ENTIRE IMAGE EDGE-TO-EDGE: no outer paper", "margin, no border area, nothing drawn outside the panels — the", "image must be mechanically cuttable along the stated cut lines.", "Panels are separated only by thin straight black gutter lines at", "the exact stated positions; ALL PANELS EXACTLY THE SAME SIZE.", ]) BLANK = "\n".join([ "There are {n_shots} shots but {slots} panel slots: draw the shots in", "panels 1..{n_shots} and leave the remaining panel(s) COMPLETELY", "BLANK PURE WHITE — empty white paper, not even a frame line inside.", ]) MOVEMENT_SYS = "\n".join([ "당신은 콘티 작가를 돕는 동선 분석가다. 씬 원문(한국어 전문)과 샷", "목록을 읽고, 각 샷에서 인물이 공간을 어떻게 통과하는지(동선)를", "판정하라. 원문에 근거한 것만 쓰고 발명 금지. 마커나 ID 코드 금지.", "샷마다 두 필드를 채워라:", "1) movement_en — 영어 한 문장: 누가/무엇이 움직이는지, 이동 경로와", "방향(오르는지 내려오는지, 들어가는지 나오는지, 카메라 쪽인지", "반대인지), 몸의 진행 자세, 어디에서 와서 어디로 가는 중인지.", "정지 샷이면 'static:'으로 시작해 시선·몸 방향만.", "★인물의 신체·의식 상태가 원문에 있으면 자세 서술에 반드시 반영:", "죽었거나 의식이 없으면 'lifeless: collapsed, asymmetric, limp", "limbs, weight fallen' 처럼 무너진 자세로 명시(단정하게 앉거나", "웅크린 자세로 쓰지 말 것), 부상·긴장·공포도 자세에 드러나게.", "어둠 속 실루엣·그림자로 보이는 존재도 실제로는 사람이면 사람의", "신체 비례임을 유지해 서술하라(예: 'a real person appearing as a", "dark silhouette, normal human build').", "2) figures_en — 영어 1~2문장: 프레임에 보이는 인물/핵심 물체", "각각에 대해 (a) 프레임 안 겉보기 크기(화면 높이의 몇 분의 몇", "수준인지: e.g. fills ~2/3 of frame height / small ~1/5), (b) 카메라", "기준 원근 서열(which is NEARER, which is FARTHER — 명시적 비교),", "(c) 인물이 여럿이거나 동선이 얽히면 서로의 경로·거리 관계까지.", "인물 1명 단순 샷이어도 크기와 근원(near/far)은 반드시 기록.", ]) MOVEMENT_SCHEMA = { "type": "object", "properties": {"movements": {"type": "array", "items": { "type": "object", "properties": {"shot": {"type": "string"}, "movement_en": {"type": "string"}, "figures_en": {"type": "string"}}, "required": ["shot", "movement_en", "figures_en"]}}}, "required": ["movements"], } def stage_movement(groups_filter): plan = F.load_plan(PLAN) mv = plan.setdefault("movement", {}) for gkey, name, _place, keys in GROUPS: if groups_filter and gkey not in groups_filter: continue scene_ids = sorted({si for si, _ in keys}) scene_txt = "\n\n".join( f"[씬 {si}] {DATA['scenes'][str(si)]['heading']}\n" + DATA['scenes'][str(si)]['text'] for si in scene_ids) shots_txt = "\n".join( f"S{si}sh{shi}: {SHOT[(si, shi)]['desc']}" for si, shi in keys) user = (f"씬 원문 전문:\n{scene_txt}\n\n샷 목록(각각 동선 판정):\n" f"{shots_txt}") res = F.llm(f"s34_move_{gkey}", MOVEMENT_SYS, user, MOVEMENT_SCHEMA) mv[gkey] = {m["shot"]: {"movement": m["movement_en"], "figures": m["figures_en"]} for m in res["movements"]} print(f"[movement] {gkey}({name}): {len(mv[gkey])}건") F.save_plan(PLAN, plan) def _prompt(place_en, keys, movement=None): n = len(keys) slots, layout, _size, _ar = _grid(n) lines = [] for i, (si, shi) in enumerate(keys, 1): s = SHOT[(si, shi)] line = f"PANEL {i}: {s['desc']}" mv = (movement or {}).get(f"S{si}sh{shi}") if mv: if isinstance(mv, str): mv = {"movement": mv, "figures": ""} line += f"\n MOVEMENT (follow exactly): {mv['movement']}" if mv.get("figures"): line += ("\n FIGURES — apparent size & depth order" f" (follow exactly): {mv['figures']}") lines.append(line) parts = [ HEAD.format(layout=layout), f"THE LOCATION (same in all drawn panels): {place_en}.", PURPOSE, REALIZE, GEOMETRY, CONSISTENCY, "PANELS (one shot per panel, Korean shot text):", "\n".join(lines), ] if slots > n: parts.append(BLANK.format(n_shots=n, slots=slots)) parts.append(SEAL.format(numbers=", ".join(str(i) for i in range(1, n + 1)))) return "\n\n".join(parts) def stage_prompts(): try: old = F.load_plan(PLAN) movement = old.get("movement", {}) except Exception: movement = {} plan = {"groups": {}, "movement": movement} for gkey, name, place, keys in GROUPS: slots, layout, size, ar = _grid(len(keys)) plan["groups"][gkey] = { "name": name, "n": len(keys), "slots": slots, "layout": layout, "i2_size": size, "nb2_ar": ar, "shots": [f"S{si}sh{shi}" for si, shi in keys], "prompt": _prompt(place, keys, movement.get(gkey)), **{eng: f"conti24_{gkey}_{eng}.png" for eng in ENGINES}, } F.save_plan(PLAN, plan) n_mv = sum(len(v) for v in movement.values()) print(f"[prompts] {len(GROUPS)}그룹 저장 (동선 주입 {n_mv}건)") def stage_gen(engines, groups_filter): plan = F.load_plan(PLAN) for gkey, _name, _place, _keys in GROUPS: if groups_filter and gkey not in groups_filter: continue e = plan["groups"][gkey] for eng in [x for x in ENGINES if x in engines]: out = OUTD / e[eng] if eng == "i2": F.img_gpt(f"s34v2_conti_{gkey}_i2", e["prompt"], refs=None, size=e["i2_size"], out_path=out) else: F.img_nb2(f"s34v2_conti_{gkey}_nb2", e["prompt"], [], aspect_ratio=e["nb2_ar"], out_path=out) print(f"[gen] {gkey} 완료") def stage_html(): plan = F.load_plan(PLAN) try: v1 = F.load_plan(PLAN_V1) except Exception: v1 = None def esc(t): return _html.escape(t or "") rows = [] for gkey, name, _place, keys in GROUPS: e = plan["groups"][gkey] shots_html = "".join( f"
  • P{i} S{si}sh{shi} — {esc(SHOT[(si, shi)]['desc'])}
    • " for i, (si, shi) in enumerate(keys, 1)) blank = (f" · 백색 패널 {e['slots'] - e['n']}개" if e['slots'] > e['n'] else "") def _img_file(eng, e=e, gkey=gkey): if (OUTD / e[eng]).exists(): return e[eng], "v24(인물=사람 형상 원칙)" for ver, fn in (("v22(EN+비유 현실화)", f"conti22_{gkey}_{eng}.png"), ("v20(실사영화·EN)", f"conti20_{gkey}_{eng}.png"), ("v9", f"conti9_{gkey}_{eng}.png"), ("v8", f"conti8_{gkey}_{eng}.png"), ("v7", f"conti7_{gkey}_{eng}.png"), ("v6", f"conti6_{gkey}_{eng}.png"), ("v5", f"conti5_{gkey}_{eng}.png"), ("v4", f"conti4_{gkey}_{eng}.png"), ("v3", f"conti3_{gkey}_{eng}.png"), ("v2", f"conti2_{gkey}_{eng}.png")): if (OUTD / fn).exists(): return fn, ver return e[eng], "생성 전" imgs = "" for eng in ENGINES: fn, ver = _img_file(eng) imgs += (f"
    " f"" f"
    {eng} · {ver}
    ") for prev, plab in ((f"conti23_{gkey}_{eng}.png", "비교 — v23(무광 검정 금지 하드룰, 반려)"), (f"conti22_{gkey}_{eng}.png", "비교 — v22(EN+비유 현실화)"), (f"conti21_{gkey}_{eng}.png", "비교 — v21(한국어 원어, 반려)"), (f"conti20_{gkey}_{eng}.png", "비교 — v20(실사영화·영어)")): if fn != prev and (OUTD / prev).exists(): imgs += (f"
    {eng} · {plab}" "
    ") if gkey == "busstop": imgs += ("
    " "
    ★참조 이미지 (현행 v5, i2=gpt-image-2 생성) — 완전 수직 2D" " 평면도(원근 0, 셸터=평면 심볼) + LLM이 샷 원문을" " 재분석한 카메라 배치(액션 뒤 down-road에서 up-road" " 방향, H 가까이·S 멀리 깊이 순서) — 오버레이도 t2i" "
    " "
    i2 · v16 (현행 최종)" " — 재판정 카메라: 혜수 등 뒤 시점, H 전신 가까이/S" " 멀리 뒷모습, 찻길 명확
    " "
    " "
    비교 — nb2 평면도 재시도: 1차보다 개선" "됐으나 ★원근 잔존(도로 사다리꼴 수렴·호안 벽 측면" " 노출)+바다·보도 실사 질감 혼입 — 완전 평면 실패" "
    " "
    " "
    비교 — nb2 오버레이(재판정 카메라):" " ★배치는 이번엔 정확(카메라 down-road·콘이 셸터+" "H+S 포함·S 화살표 교차로 방향·단일성 준수) — 단" " 바탕 맵의 원근·실사 혼입 상속
    " "
    i2 · v17 — nb2" " 오버레이 맵 기반: 재판정 카메라·깊이 순서·찻길 전부" " 정확 (바탕 맵의 원근·실사 혼입이 출력에 전염되지" " 않음)
    " "
    nb2 · v17 — 같은" " 맵 기반: 배치·찻길 정확+텍스트 무재발, 단 등분 위반" " 지속(상단 몰림·하단 여백)
    " "
    i2 · v18 (재현 1)" " — 동일 구도 유지
    " "
    nb2 · v18 (재현 1)" " — ★등분 첫 준수·배치 정확 (단 밤이 낮처럼 밝음)" "
    " "
    i2 · v19 (재현 2)" " — 동일 구도 유지
    " "
    nb2 · v19 (재현 2)" " — 배치 정확·등분 재위반
    " "
    " "
    이력 — v4 참조(구 카메라 판정·준평면)" "
    " "
    구판 맵 (4회차 캐논 —" " 도로 소실 결함 실측, 비교용)
    " "
    i2 · v15 (현행 최종)" " — t2i 주석판 맵 참조: 찻길+보도+셸터+호안 층 구조" " 정확
    " "
    비교: nb2 맵 재투영" " — 탑다운 미준수(아이소메트릭)+하단 실사 질감 혼입" "+B2 마커 누락
    " "
    " "
    비교: nb2 오버레이 — 단일성(1H/1S/1카메라)" "은 준수했으나 ★시야 콘이 지시 반대 방향으로 열림" "(S가 콘 밖)+S가 차도 위
    ") for eng2 in ENGINES: for fn_x, lab in ( (f"conti12_busstop_{eng2}.png", "맵+실사+카메라 grounding 참조 (화살표 無)"), (f"conti13_busstop_{eng2}.png", "맵+카메라 grounding만 — 실사 제외 ablation")): if (OUTD / fn_x).exists(): imgs += (f"
    {eng2} · " f"{lab} — P2 전용
    ") v1imgs = "" if v1 and gkey in v1.get("groups", {}): v1e = v1["groups"][gkey] v1imgs = ("
    v1 (실사판 — 비교용)
    " + "".join( f"
    " f"" f"
    v1 {eng}
    " for eng in ENGINES if eng in v1e) + "
    ") rows.append( f"

    {esc(name)} ({gkey} · 샷 {e['n']} → " f"{e['slots']}분할{blank})

    {imgs}
    " f"
    패널 구성 (샷 원문)
      {shots_html}" f"
    {esc(e['prompt'])}
    {v1imgs}
    ") still_section = "" if (OUTD / "still_S15sh6_i2.png").exists(): figs = "" for fn, cap in ( ("panel_S15sh2_nb2.png", "크롭 패널 P1 (nb2 v18 콘티 좌반)"), ("panel_S15sh6_nb2.png", "크롭 패널 P2 (우반)"), ("still_S15sh2_i2.png", "스틸 S15sh2 · i2"), ("still_S15sh2_nb2.png", "스틸 S15sh2 · nb2"), ("still_S15sh6_i2.png", "스틸 S15sh6 · i2 (prev 없음)"), ("still_S15sh6_nb2.png", "스틸 S15sh6 · nb2 (prev 없음)"), ("still2_S15sh6_i2.png", "스틸 S15sh6 · i2 + prev 동반 — 장소 연속 개선, ★인물 역할" " 스왑(남은 이가 수리영 의상)"), ("still2_S15sh6_nb2.png", "스틸 S15sh6 · nb2 + prev 동반 — ★의상·장소 연속 정확" "(라벨드 참조가 역할 고정)"), ("still3_S15sh6_i2.png", "스틸 S15sh6 · i2 + prev + 역할 바인딩 명시 — 그래도 스왑" " 잔재(2연속): i2 다중 인물 참조 정체성 관리 약함"), ): if (OUTD / fn).exists(): figs += (f"
    " f"" f"
    {fn.split('_')[0]} — {cap}" "
    ") still_section = ( "

    정류장 — 콘티 패널 → 실사 스틸 체인 " "(nb2 v18 콘티를 코드로 반분 → 패널=구도 참조 + " "캐릭터 합성 참조 2장만(배경 참조 제외) → i2/nb2 스틸)" f"

    {figs}
    ") PAGE.write_text(f""" s34 콘티 시트 v2 — 러프 스케치 · 연속성 6그룹 × i2/nb2

    s34 v2 — 손그림 러프 스케치 콘티 (구도·연속성 전용, 균등 분할)

    계약: 손그림 연필 러프 스케치(실사 금지)·글자 절대 금지(샷 번호만)· 2/4/6 균등 분할·패널 수=샷 수, 홀수면 잔여 패널 순백(숲 3샷→2×2+백색 1, 경찰서 5샷→3×2+백색 1)·모션 화살표로 인물/카메라 이동 표기·전 패널 같은 장소 연속성. 균등 그리드라 코드가 패널을 잘라 분해 가능. i2/nb2 둘 다.
    {still_section} {''.join(rows)} """) print(f"[html] {PAGE}") CANON_DIR = Path("/Users/manta/Documents/Projects/TheRoad-I1/projects/" "8207aadc-7975-48f5-af17-ccc145a6660d/images/background_chain") BUS_PHOTO = CANON_DIR / "canon_bg_coastal_bus_stop_master.png" BUS_MAP = CANON_DIR / "canon_bg_coastal_bus_stop_map.png" BUS_P2_GROUND = "\n".join([ "PANEL 2 CAMERA & PLACEMENT (follow exactly, from the site plan):", "camera stands on the water-side portion of the bus stop, just", "beyond the bench, looking along the length of the coastal road —", "in frame: the departing runner in full-body stride at center,", "the bench with the other woman beside it in the left background,", "the coastal land edge to the right and the open sea beyond it.", ]) BUS_REF_NOTE = "\n".join([ "REFERENCES (for grounding, not for style): the attached LOCATION", "PHOTOGRAPH shows how this exact bus stop looks (shelter, bench,", "road, sea side); the attached flat SITE PLAN shows the layout from", "above — never draw the plan itself or its markers. APPLY THEM TO", "PANEL 2 ONLY: panel 2's camera position, shelter/bench placement", "and figure positions must match the plan and the photograph's", "geography. PANEL 1 is an earlier, tighter moment from a different", "angle — take only the stop's general look (bench, shelter) from", "the references, not their viewpoint. The OUTPUT stays a rough", "hand-drawn pencil storyboard sketch — do not copy the photograph's", "rendering, only its spatial truth.", ]) BUS_MAPONLY_NOTE = "\n".join([ "REFERENCE (for grounding, not for style): the attached flat SITE", "PLAN shows this bus stop's layout from above — where the shelter,", "bench, road, land edge and sea are relative to each other. Never", "draw the plan itself or its markers. APPLY IT TO PANEL 2 ONLY:", "panel 2's camera position, shelter/bench placement and figure", "positions must match this plan. PANEL 1 is an earlier, tighter", "moment from a different angle — do not force the plan's viewpoint", "onto it. The OUTPUT stays a rough hand-drawn pencil storyboard", "sketch.", ]) def stage_busstop_maponly(engines): plan = F.load_plan(PLAN) e = plan["groups"]["busstop"] prompt = e["prompt"] + "\n\n" + BUS_P2_GROUND + "\n\n" + BUS_MAPONLY_NOTE if "i2" in engines: out = OUTD / "conti13_busstop_i2.png" F.img_gpt("s34v13_busstop_maponly_i2", prompt, refs=[BUS_MAP], size="1536x1024", out_path=out) print(f"[busstop_maponly] {out}") if "nb2" in engines: out = OUTD / "conti13_busstop_nb2.png" F.img_nb2("s34v13_busstop_maponly_nb2", prompt, [("SITE PLAN — top-down layout source only; never draw" " this plan or its markers.", BUS_MAP)], aspect_ratio="16:9", out_path=out) print(f"[busstop_maponly] {out}") def stage_busstop_refs(engines): plan = F.load_plan(PLAN) e = plan["groups"]["busstop"] prompt = e["prompt"] + "\n\n" + BUS_P2_GROUND + "\n\n" + BUS_REF_NOTE if "i2" in engines: out = OUTD / "conti12_busstop_i2.png" F.img_gpt("s34v12_busstop_refs_i2", prompt, refs=[BUS_PHOTO, BUS_MAP], size="1536x1024", out_path=out) print(f"[busstop_refs] {out}") if "nb2" in engines: out = OUTD / "conti12_busstop_nb2.png" refs = [("LOCATION PHOTOGRAPH — this exact bus stop; spatial truth" " only, output stays a pencil sketch.", BUS_PHOTO), ("SITE PLAN — top-down layout source only; never draw this" " plan or its markers.", BUS_MAP)] F.img_nb2("s34v12_busstop_refs_nb2", prompt, refs, aspect_ratio="16:9", out_path=out) print(f"[busstop_refs] {out}") if __name__ == "__main__": ap = argparse.ArgumentParser() ap.add_argument("--only", required=True, choices=["prompts", "movement", "gen", "html", "busstop_refs", "busstop_maponly"]) ap.add_argument("--engines", default="i2,nb2") ap.add_argument("--groups", default="") a = ap.parse_args() if a.only == "prompts": stage_prompts() elif a.only == "busstop_refs": stage_busstop_refs(a.engines.split(",")) elif a.only == "busstop_maponly": stage_busstop_maponly(a.engines.split(",")) elif a.only == "movement": stage_movement([g for g in a.groups.split(",") if g]) elif a.only == "gen": stage_gen(a.engines.split(","), [g for g in a.groups.split(",") if g]) else: stage_html()