dolphin_vision_director.py

# -*- coding: utf-8 -*-
import os
import re
import io
import json
import base64
import hashlib
import sqlite3
import urllib.request
import urllib.error

import numpy as np
import folder_paths
import comfy.sd
import comfy.lora
import comfy.utils
from safetensors import safe_open
from comfy_execution.graph_utils import ExecutionBlocker

try:
    from PIL import Image
    _HAS_PIL = True
except Exception:
    _HAS_PIL = False

CACHE_FILE = os.path.join(os.path.dirname(__file__), "dolphin_tags_cache.json")


# ======================================================================
#  모델 프리셋 (텍스트 LLM = 4샷 분배 / 비전 VLM = 이미지 인식)
# ======================================================================
LLM_PRESETS = {
    "🔞 DeepSeek V3.2 (무삭제·JSON안정·추천)": "deepseek/deepseek-v3.2",
    "🔞 Magnum v4 72B (표현 자연스러움)": "anthracite-org/magnum-v4-72b",
    "🔞 Cydonia 24B v4.1 (경량·빠름)": "thedrummer/cydonia-24b-v4.1",
    "🔞 Dolphin Mistral Venice (무료)": "cognitivecomputations/dolphin-mistral-24b-venice-edition:free",
    "⚙️ custom (아래 텍스트 사용)": "__custom__",
}
LLM_PRESET_LABELS = list(LLM_PRESETS.keys())

# 비전 모델: 특수 상황 이해 + 다국어. Qwen2.5-VL 계열이 포즈/맥락 파악에 강함.
VLM_PRESETS = {
    "👁 Qwen2.5-VL 72B (특수상황·추천)": "qwen/qwen2.5-vl-72b-instruct",
    "👁 Qwen2.5-VL 32B (균형)": "qwen/qwen2.5-vl-32b-instruct",
    "👁 Qwen2.5-VL 32B (무료)": "qwen/qwen2.5-vl-32b-instruct:free",
    "👁 Mistral Small 3.1 24B (Pixtral)": "mistralai/mistral-small-3.1-24b-instruct",
    "👁 Gemma 3 27B (경량)": "google/gemma-3-27b-it",
    "⚙️ custom (아래 텍스트 사용)": "__custom__",
}
VLM_PRESET_LABELS = list(VLM_PRESETS.keys())


def resolve_slug(table, label, custom_text, default):
    slug = table.get(label, "__custom__")
    if slug == "__custom__":
        return custom_text.strip() or default
    return slug


def resolve_api_key(widget_value):
    """위젯에 키가 있으면 그대로 사용. 비어있으면 OPENROUTER_API_KEY 환경변수로 폴백.
    -> 워크플로우 JSON(공유/백업 시 유출 위험)에 실키를 박아둘 필요가 없어짐."""
    key = (widget_value or "").strip()
    if key:
        return key
    return os.environ.get("OPENROUTER_API_KEY", "").strip()


# ======================================================================
#  캐시 & 트리거 (검증된 로직 재사용)
# ======================================================================
def _load_cache():
    if os.path.exists(CACHE_FILE):
        try:
            with open(CACHE_FILE, "r", encoding="utf-8") as f:
                return json.load(f)
        except (json.JSONDecodeError, OSError):
            return {}
    return {}


def _save_cache(cache):
    try:
        with open(CACHE_FILE, "w", encoding="utf-8") as f:
            json.dump(cache, f, indent=4, ensure_ascii=False)
    except OSError:
        pass


def _file_sig(path):
    st = os.stat(path)
    return f"{os.path.basename(path)}::{st.st_size}::{int(st.st_mtime)}"


def _sha256(path):
    h = hashlib.sha256()
    with open(path, "rb") as f:
        for block in iter(lambda: f.read(4096 * 1024), b""):
            h.update(block)
    return h.hexdigest()


# CivitAI 및 미러(civitaired) 해시 조회 엔드포인트. 앞에서부터 순서대로 시도.
TRIGGER_HASH_APIS = [
    "https://civitai.com/api/v1/model-versions/by-hash/{h}",
    "https://civitaired.com/api/v1/model-versions/by-hash/{h}",
]

# LoRA Manager(comfyui-lora-manager)가 스캔 시 만들어두는 로컬 SQLite 캐시.
# sha256으로 trained_words를 즉시 조회 가능 -> 네트워크 호출 없이 트리거워드 확보.
LORA_MANAGER_DB = os.path.join(
    os.environ.get("LOCALAPPDATA", ""), "ComfyUI-LoRA-Manager", "cache", "model", "comfyui.sqlite")


def _query_lora_manager_db(digest):
    """LoRA Manager SQLite 캐시에서 sha256으로 trained_words 조회. 실패/미발견 시 None."""
    if not digest or not os.path.exists(LORA_MANAGER_DB):
        return None
    try:
        uri = f"file:{LORA_MANAGER_DB}?mode=ro"
        con = sqlite3.connect(uri, uri=True, timeout=3)
        try:
            cur = con.cursor()
            cur.execute(
                "SELECT trained_words FROM models WHERE model_type='lora' AND sha256=? LIMIT 1",
                (digest,))
            row = cur.fetchone()
        finally:
            con.close()
        if not row or not row[0]:
            return None
        tw = json.loads(row[0])
        return tw if isinstance(tw, list) else None
    except (sqlite3.Error, OSError, ValueError, json.JSONDecodeError) as e:
        print(f"⚠️ [Director] LoRA Manager DB 조회 실패: {type(e).__name__}: {e}")
        return None


def _query_trigger_api(url):
    """단일 해시 조회 엔드포인트에서 trainedWords 리스트를 반환. 실패 시 ([], 사유)."""
    try:
        req = urllib.request.Request(url, headers={'User-Agent': 'Mozilla/5.0'})
        with urllib.request.urlopen(req, timeout=8) as r:
            data = json.loads(r.read().decode('utf-8'))
            return (data.get("trainedWords", []) or []), None
    except (urllib.error.URLError, urllib.error.HTTPError, TimeoutError,
            OSError, ValueError, json.JSONDecodeError) as e:
        return [], f"{type(e).__name__}: {e}"


def get_triggers(lora_name, cache):
    if lora_name == "None":
        return []
    path = folder_paths.get_full_path("loras", lora_name)
    if not path or not os.path.exists(path):
        return []
    try:
        with safe_open(path, framework="pt", device="cpu") as f:
            m = f.metadata()
            if m and "modelspec.trigger_words" in m:
                tw = json.loads(m["modelspec.trigger_words"])
                if isinstance(tw, list):
                    return tw
                if isinstance(tw, str):
                    return [tw]
    except (OSError, ValueError, json.JSONDecodeError):
        pass
    try:
        sig = _file_sig(path)
    except OSError:
        sig = lora_name
    # 캐시에 '비어있지 않은' 결과가 있을 때만 신뢰. 과거에 실패로 저장된 빈 값은
    # 무시하고 재조회한다 (빈 값 영구화 회귀 방지).
    cached = cache.get(sig)
    if cached:
        return cached

    # 해시는 1회만 계산해서 LoRA Manager DB -> CivitAI -> civitaired 순으로 조회.
    result, base = [], os.path.basename(path)
    try:
        digest = _sha256(path)
    except OSError as e:
        print(f"⚠️ [Director] 해시 계산 실패({base}): {e}")
        digest = None
    if digest:
        found = _query_lora_manager_db(digest)
        if found:
            print(f"✅ [Director] 트리거 조회 성공({base}) @ LoRA Manager DB: {found}")
            result = found
    if digest and not result:
        for tmpl in TRIGGER_HASH_APIS:
            host = tmpl.split('/')[2]
            found, err = _query_trigger_api(tmpl.format(h=digest))
            if found:
                print(f"✅ [Director] 트리거 조회 성공({base}) @ {host}: {found}")
                result = found
                break
            print(f"⚠️ [Director] 트리거 조회 실패({base}) @ {host}: {err or 'no trainedWords'}")

    # 성공(비어있지 않음)일 때만 캐싱. 빈 결과는 저장하지 않아 다음 실행에서 재시도됨.
    if result:
        cache[sig] = result
        _save_cache(cache)
    return result


def apply_model_lora(model, lora_data, weight):
    try:
        return comfy.sd.load_lora_for_models(model, None, lora_data, weight, 0)[0]
    except (AttributeError, TypeError):
        new_model = model.clone()
        key_map = comfy.lora.model_lora_keys_unet(new_model.model)
        loaded = comfy.lora.load_lora(lora_data, key_map)
        new_model.add_patches(loaded, weight)
        return new_model


def tensor_to_base64(image_tensor, max_side=1024):
    """ComfyUI IMAGE 텐서(B,H,W,C, 0~1 float) 첫 프레임을 JPEG base64 data URI로."""
    if not _HAS_PIL:
        print("❌ [Director] PIL(Pillow) 없음 - 비전 불가. `pip install Pillow` 필요.")
        return None
    if image_tensor is None:
        print("❌ [Director] image 입력이 None - IMAGE 연결 확인.")
        return None
    try:
        img = image_tensor
        if hasattr(img, "cpu"):
            img = img.cpu().numpy()
        img = np.asarray(img)
        if img.ndim == 4:
            img = img[0]
        arr = np.clip(img * 255.0, 0, 255).astype(np.uint8)
        pil = Image.fromarray(arr)
        w, h = pil.size
        if max(w, h) > max_side:
            scale = max_side / float(max(w, h))
            pil = pil.resize((int(w * scale), int(h * scale)), Image.LANCZOS)
        buf = io.BytesIO()
        pil.convert("RGB").save(buf, format="JPEG", quality=90)
        b64 = base64.b64encode(buf.getvalue()).decode("utf-8")
        print(f"🖼 [Director] image encoded ok: {w}x{h} -> {pil.size}, {len(b64)//1024}KB base64")
        return f"data:image/jpeg;base64,{b64}"
    except Exception as e:
        print(f"❌ [Director] image encode failed: {type(e).__name__}: {e}")
        return None


# ======================================================================
#  Vision Director: 이미지 인식 + 필수동작 실행 + 클립당 4로라
# ======================================================================
class DolphinVisionDirector:
    """
    이미지 하나와 '필수 동작'과 로라만 넣으면 4개 연속 클립이 완성되는 올인원 노드.

      1) IMAGE -> 비전 VLM 이 상황(캐릭터 상태/자세/환경)만 앵커로 파악
      2) 그 맥락 위에서 '내가 준 필수 동작(essential_actions)'만 4샷으로 분배
      3) 클립마다 로라 4개씩 개별 적용(웨이트도 개별) + 트리거 자동 주입
    """

    @classmethod
    def INPUT_TYPES(s):
        loras = ["None"] + (folder_paths.get_filename_list("loras") or [])
        req = {
            "image": ("IMAGE",),
            "model_base": ("MODEL",),
            "model_high": ("MODEL",),
            "model_low": ("MODEL",),
            "essential_actions": ("STRING", {"multiline": True,
                "default": ("돌진한다, 검을 휘둘러 적을 쓰러뜨린다, 총알을 튕겨낸다, 숨통을 끊는다")}),
            "num_clips": (["4 (20s)", "3 (15s)", "2 (10s)"], {"default": "4 (20s)",
                "tooltip": "실제로 렌더링할 클립 수. Extend range(Fast Groups Bypasser) 토글과 맞춰서 "
                           "설정해야 스토리가 그 길이에 맞게 배분됨. 나머지 클립 슬롯은 마지막 클립 내용을 반복."}),
            "character_name": ("STRING", {"default": "AUTO"}),
            "artistic_vibe": ("STRING", {"multiline": True,
                                         "default": "cinematic lighting, dynamic motion"}),
            "pacing": (["auto", "slow-burn (1 beat)", "balanced (multi beat)", "rapid (4 distinct)"],
                       {"default": "auto"}),
            "detail_level": (["짧게(간결)", "필수+앵커만", "상세"], {"default": "필수+앵커만"}),
            "inject_triggers": ("BOOLEAN", {"default": True,
                                            "label_on": "TRIGGERS ON", "label_off": "TRIGGERS OFF"}),
            "use_vision": ("BOOLEAN", {"default": True,
                                       "label_on": "VISION ON", "label_off": "VISION OFF (text only)"}),
            "openrouter_api_key": ("STRING", {"default": ""}),
            "vlm_preset": (VLM_PRESET_LABELS, {"default": "👁 Qwen2.5-VL 72B (특수상황·추천)"}),
            "vlm_model": ("STRING", {"default": "qwen/qwen2.5-vl-72b-instruct"}),
            "llm_preset": (LLM_PRESET_LABELS, {"default": "🔞 DeepSeek V3.2 (무삭제·JSON안정·추천)"}),
            "llm_model": ("STRING", {"default": "deepseek/deepseek-v3.2"}),
            "creativity": ("FLOAT", {"default": 0.7, "min": 0.1, "max": 1.5, "step": 0.05}),
            "seed": ("INT", {"default": 0, "min": 0, "max": 0xffffffffffffffff}),
        }
        opt = {}
        for c in range(1, 5):
            for n in range(1, 5):
                opt[f"c{c}_lora_{n}"] = (loras, {"default": "None"})
                opt[f"c{c}_w_{n}"] = ("STRING", {"default": "1.0, 1.0, 1.0", "multiline": False})
        opt["manual_triggers"] = ("STRING", {"default": "", "multiline": True})
        opt["vision_caption"] = ("STRING", {"multiline": True, "default": "",
                                            "forceInput": True})
        return {"required": req, "optional": opt}

    RETURN_TYPES = (
        "MODEL", "MODEL", "MODEL", "STRING",
        "MODEL", "MODEL", "MODEL", "STRING",
        "MODEL", "MODEL", "MODEL", "STRING",
        "MODEL", "MODEL", "MODEL", "STRING",
        "STRING", "STRING", "STRING",
        "BOOLEAN", "BOOLEAN", "BOOLEAN", "BOOLEAN",
    )
    RETURN_NAMES = (
        "c1_base", "c1_high", "c1_low", "c1_prompt",
        "c2_base", "c2_high", "c2_low", "c2_prompt",
        "c3_base", "c3_high", "c3_low", "c3_prompt",
        "c4_base", "c4_high", "c4_low", "c4_prompt",
        "triggers", "scene_analysis", "debug_info",
        "c1_is_final", "c2_is_final", "c3_is_final", "c4_is_final",
    )
    FUNCTION = "direct"
    CATEGORY = "Dolphin"

    def _build_final(self, action, master_scene, name, trigger_str):
        head = ", ".join([p for p in [name, trigger_str] if p])
        body = " ".join([p for p in [master_scene.strip().strip(",. "), action.strip()] if p])
        if head and body:
            return f"{head}\n{body}"
        return head or body

    def _pacing_hint(self, p):
        return {
            "auto": "Decide pacing yourself based on how many distinct beats the actions contain.",
            "slow-burn (1 beat)": "One climactic beat: clips 1-3 build tension, clip 4 detonates.",
            "balanced (multi beat)": "Early clips build, later clips deliver; keep motion continuous.",
            "rapid (4 distinct)": "Four distinct consecutive actions, one per clip, each kinetic.",
        }.get(p, "Decide pacing yourself.")

    def _detail_hint(self, d):
        return {
            "짧게(간결)": "Keep each clip to a short punchy motion phrase. Pure movement verbs only.",
            "필수+앵커만": "Each clip is one concise continuous motion. No description, only movement.",
            "상세": "Motion may be described in more detail, but ONLY body movement — no scenery.",
        }.get(d, "Motion only, no description.")

    def _extract_json(self, text):
        if not text:
            return None
        text = re.sub(r'```(?:json)?', '', text).strip()
        a, b = text.find('{'), text.rfind('}')
        if a == -1 or b <= a:
            return None
        blob = text[a:b + 1]
        try:
            return json.loads(blob)
        except json.JSONDecodeError:
            try:
                return json.loads(re.sub(r',\s*([}\]])', r'\1', blob))
            except json.JSONDecodeError:
                return None

    def _call_or(self, messages, model, key, temp, timeout=120, force_json=False):
        payload = {"model": model.strip(), "messages": messages, "temperature": temp}
        if force_json:
            payload["response_format"] = {"type": "json_object"}
        req = urllib.request.Request(
            "https://openrouter.ai/api/v1/chat/completions",
            data=json.dumps(payload).encode('utf-8'),
            headers={'Authorization': f'Bearer {key}', 'Content-Type': 'application/json'})
        raw = json.loads(urllib.request.urlopen(req, timeout=timeout).read().decode('utf-8'))
        if "choices" not in raw:
            err = raw.get("error") or raw
            msg = err.get("message") if isinstance(err, dict) else str(err)
            code = err.get("code") if isinstance(err, dict) else "?"
            raise ValueError(f"API error (code={code}): {msg}")
        return raw['choices'][0]['message']['content'].strip()

    def _vision_analyze(self, data_uri, vlm_model, key, temp):
        sys = ("You are a visual scene analyst for a video pipeline. "
               "Look at the image and extract ONLY concrete anchors, no storytelling. "
               "Report exactly these lines:\n"
               "(1) character appearance & clothing.\n"
               "(2) current pose/posture (sitting, crouched, standing, balance, limb positions).\n"
               "(3) SUPPORT & CONTACT: what the character is on top of, sitting on, holding, "
               "leaning against, or touching (e.g. 'seated balanced on a yoga ball', "
               "'gripping a sword', 'kneeling on the floor'). This is critical for physics — "
               "state it explicitly even if obvious.\n"
               "(4) environment & lighting (brief).\n"
               "Be terse, one short line each. English only. "
               "Do not invent actions or narrative. If content is mature, describe it plainly and factually.")
        messages = [
            {"role": "system", "content": sys},
            {"role": "user", "content": [
                {"type": "text", "text": "Extract the anchors from this image."},
                {"type": "image_url", "image_url": {"url": data_uri}},
            ]},
        ]
        import time
        last_err = "vision error"
        for attempt in range(1, 4):
            try:
                resp = self._call_or(messages, vlm_model, key, temp, timeout=90).strip()
                if not resp:
                    print(f"⚠️ [Director] VLM({vlm_model}) 응답이 비었음 (거부 or 빈 반환)")
                    return "", "vision empty response"
                print(f"👁 [Director] VLM({vlm_model}) anchors (attempt {attempt}):\n{resp[:300]}")
                return resp, None
            except urllib.error.HTTPError as e:
                body = ""
                try:
                    body = e.read().decode('utf-8')[:200]
                except Exception:
                    pass
                print(f"❌ [Director] VLM HTTP {e.code} (attempt {attempt}): {body}")
                last_err = f"vision HTTP {e.code}"
                if e.code in (429, 503) and attempt < 3:
                    wait = 2 ** attempt
                    print(f"⏳ [Director] rate limited, {wait}s 후 재시도...")
                    time.sleep(wait)
                    continue
                return "", last_err
            except (urllib.error.URLError, TimeoutError, KeyError, ValueError) as e:
                print(f"❌ [Director] VLM error (attempt {attempt}): {type(e).__name__}: {e}")
                last_err = f"vision error: {e}"
                if attempt < 3:
                    time.sleep(2 ** attempt)
                    continue
                return "", last_err
        return "", last_err

    def _fallback_name(self, anchors):
        import random
        a = (anchors or "").lower()
        east = ["asian", "east asian", "japanese", "korean", "chinese", "anime",
                "kimono", "hanbok", "qipao", "oriental"]
        western = ["western", "european", "caucasian", "american", "blonde", "redhead"]
        east_names = ["Yuna", "Kaede", "Jin", "Mei", "Haruka", "Rin", "Sora", "Aoi"]
        west_names = ["Elena", "Ryan", "Ava", "Lucas", "Mila", "Ethan", "Nora", "Leo"]
        if any(k in a for k in east):
            pool = east_names
        elif any(k in a for k in western):
            pool = west_names
        else:
            pool = east_names + west_names
        return random.choice(pool)

    def _split_actions_to_clips(self, actions, num_clips=4):
        parts = re.split(r'(?<=[.!?。])\s+|[,、]|\n+', actions)
        parts = [p.strip() for p in parts if len(p.strip()) > 1]
        if not parts:
            parts = [actions.strip()]
        n = len(parts)
        last = num_clips
        assign = {c: [] for c in range(1, num_clips + 1)}
        if num_clips == 1:
            assign[1] = parts
            return assign
        build_labels = ["(build-up toward)", "(begin)", "(escalate)", "(continue escalating)"]
        if n == 1:
            for c in range(1, last):
                label = build_labels[min(c - 1, len(build_labels) - 1)]
                assign[c] = [f"{label} {parts[0]}"]
            assign[last] = [parts[0]]
        elif n <= num_clips:
            assign[last] = [parts[-1]]
            head = parts[:-1]
            for i, ph in enumerate(head):
                assign[min(i + 1, last - 1)].append(ph)
            for c in range(1, last):
                if not assign[c]:
                    nxt = parts[min(c, n - 1)]
                    assign[c] = [f"(move toward) {nxt}"]
        else:
            assign[last] = [parts[-1]]
            rest = parts[:-1]
            k, m = divmod(len(rest), last - 1)
            idx = 0
            for c in range(1, last):
                end = idx + k + (1 if (c - 1) < m else 0)
                assign[c] = rest[idx:end]
                idx = end
        return assign

    def _plan_shots(self, anchors, actions, name, vibe, pacing, detail, llm_model, key, temp, seed, num_clips=4):
        assign = self._split_actions_to_clips(actions, num_clips)
        assign_block = "\n".join(
            f"  clip_{c}: {', '.join(assign[c]) if assign[c] else '(continue previous motion)'}"
            for c in range(1, num_clips + 1))

        last_clip = num_clips
        total_seconds = num_clips * 5
        if num_clips >= 3:
            chain = " and ".join(f"clip_{c}<-clip_{c-1}" for c in range(3, num_clips + 1))
            continuity_middle = (f". The same applies {chain}: each clip picks up exactly where the "
                                  f"previous clip's body position left off")
        else:
            continuity_middle = ""
            
        continuity_rule = ""
        if num_clips >= 2:
            continuity_rule = (
                "3) INTER-CLIP CONTINUITY. clip_2 MUST begin from the exact physical end-state of "
                "clip_1's action (same position, body orientation, and momentum) — it continues the "
                "motion, it does not restart from a neutral stance" + continuity_middle +
                f", as one unbroken {total_seconds}-second performance rather than {num_clips} separate poses. Do "
                "NOT explicitly narrate the hand-off (no phrases like 'continuing from', 'still', 'as "
                "before', 'picking up where') — simply choreograph the new action so it is kinematically "
                "consistent with how the previous action would have left the body positioned. This is a "
                "physical-continuity constraint, not a scripted transition line.\n"
            )

        json_clip_fields = ",".join(
            f'"clip_{i}":"<only clip_{i} action{", the finisher" if i == last_clip else ""}, '
            f'continuous & fast>"' for i in range(1, num_clips + 1))

        sys = (
            f"You are an elite fight/action choreographer for a {total_seconds}-second video made of "
            f"{num_clips} consecutive 5-second clip{'s' if num_clips != 1 else ''}. You are given SCENE ANCHORS "
            "(the character's starting look/pose/place from an image) and a STRICT per-clip action "
            "assignment.\n"
            "\n"
            "ABSOLUTE RULES:\n"
            "1) FOLLOW THE ASSIGNMENT EXACTLY. Each clip animates ONLY the action(s) assigned "
            "to it below. You MUST NOT move a later clip's action into an earlier clip. "
            f"clip_1 must NOT contain clip_{last_clip}'s action. The final action happens ONLY in "
            f"clip_{last_clip}. "
            "This is the most important rule - violating the order is a failure. "
            "(ONE exception: clip_1 may prepend a brief transition OUT of the starting physics from "
            "rule 2 — e.g. leaving the yoga ball — before its assigned action. This is not borrowing "
            "a later action; it is grounding the first one.)\n"
            "2) STARTING PHYSICS (use the anchors here). Read the character's CURRENT pose AND "
            "what they are supported by / in contact with from the SCENE ANCHORS (e.g. seated on a "
            "yoga ball, kneeling, gripping a weapon, leaning on a wall). clip_1 MUST begin the first "
            "action FROM that exact physical situation and honor its constraints. If the support is "
            "unstable or unusual (a yoga ball, a ledge, a moving surface), the motion MUST account "
            "for it — e.g. anchor 'seated balanced on a yoga ball' -> clip_1 = 'pushing off the "
            "wobbling ball and rising to their feet as it rolls away' BEFORE any further action. "
            "Never ignore or contradict the starting support (do not stand if seated on the ball "
            "without first leaving it; do not assume a weapon is drawn if the anchor shows it "
            "sheathed). This makes the motion physically continuous with the input image. Reference "
            "the support ONLY to ground how the motion STARTS — do NOT describe its appearance, "
            "color, or the wider environment.\n"
            + continuity_rule +
            "4) ACTION ONLY. Write physical body movement. NO scenery, NO lighting, NO clothing, "
            "NO mood, NO camera talk. If a word is not a movement or body part, delete it. "
            "(Referring to the starting pose in clip_1 per rule 2, or to the prior clip's ending "
            "position per rule 3, is allowed since it describes how the body moves.)\n"
            "5) FILL 5 SECONDS. Expand the assigned action into a single continuous flow of motion "
            "that occupies the whole clip (use 'then', 'immediately', 'without pausing'). "
            "Do not borrow the next action to fill time - elaborate the CURRENT action instead.\n"
            "6) ANTI-SLOW-MOTION: real-time speed, kinetic adverbs (swiftly, rapidly, explosively, "
            "in a split second). NO slow motion, NO freeze, NO holding a pose.\n"
            "7) CHARACTER NAME: study the SCENE ANCHORS. Judge the character's apparent ethnicity/"
            "setting (East Asian, Western, etc.) and INVENT a specific fitting first name that "
            "matches it (e.g. East Asian -> 'Yuna','Kaede','Jin'; Western -> 'Elena','Ryan'). "
            "NEVER output 'AUTO' or an empty name. If a NAME is given below (not 'AUTO'), use it.\n"
            f"8) {detail}\n"
            "9) MOTION ENERGY: a MOTION STYLE cue is given below (e.g. 'cinematic lighting, dynamic "
            "motion'). Use it ONLY to calibrate how forceful/graceful/frantic the movement FEELS "
            "(word choice, verb intensity). Do NOT quote it and do NOT let any scenery/lighting/mood "
            "words from it leak into the output — rule 4 still applies.\n"
            "\n"
            "master_scene: leave it EMPTY.\n"
            "\n"
            'OUTPUT ONLY JSON: {"character":"<a real name, never AUTO>","master_scene":"",'
            + json_clip_fields + "}"
        )
        usr = (f"NAME: {name}\nPACING: {self._pacing_hint(pacing)}\n\n"
               f"MOTION STYLE (kinetic energy cue only, per rule 9 — do not describe scenery/"
               f"lighting/mood from this): {vibe.strip() or '(none)'}\n\n"
               f"SCENE ANCHORS (use the POSE to start clip_1's motion per rule 2; "
               f"use overall look only for naming — do NOT copy look into the clips):\n"
               f"{anchors or '(none — assume a neutral ready stance)'}\n\n"
               f"STRICT PER-CLIP ACTION ASSIGNMENT (animate ONLY what is listed per clip, "
               f"in this order, finisher in clip_{last_clip}):\n{assign_block}\n")
        messages = [{"role": "system", "content": sys}, {"role": "user", "content": usr}]
        t = temp
        last = "⚠️ plan failed (no attempts ran)"
        for attempt in (1, 2):
            try:
                content = self._call_or(messages, llm_model, key, t, force_json=True)
                parsed = self._extract_json(content)
                if parsed and all(parsed.get(f"clip_{i}") for i in range(1, num_clips + 1)):
                    return parsed, f"✅ plan ok (attempt {attempt})"
                last = f"⚠️ incomplete JSON (attempt {attempt}): {content[:150]!r}"
                print(f"⚠️ [Director] plan attempt {attempt} returned incomplete JSON: {content[:300]}")
            except urllib.error.HTTPError as e:
                body = ""
                try:
                    body = e.read().decode('utf-8')[:200]
                except Exception:
                    pass
                last = f"❌ plan HTTP {e.code} (attempt {attempt}): {body}"
                print(f"❌ [Director] plan {last}")
            except (urllib.error.URLError, TimeoutError, KeyError, ValueError) as e:
                last = f"❌ plan error (attempt {attempt}): {type(e).__name__}: {e}"
                print(f"❌ [Director] {last}")
            t = min(temp, 0.4)
        return None, last

    def _local_split(self, actions, num_clips=4):
        parts = re.split(r'(?<=[.!?。])\s+|[,、]|\n+', actions)
        parts = [p.strip() for p in parts if len(p.strip()) > 1]
        if not parts:
            parts = [actions.strip()]
        n = len(parts)
        last = num_clips - 1

        def fast(phrase, lead="swiftly"):
            p = phrase.strip()
            return f"{lead} {p}, continuous real-time motion, no slow motion" if p else ""

        if num_clips == 1:
            return [fast(", ".join(parts), "explosively")]

        if n == 1:
            a = parts[0]
            templates = [
                f"rapidly moves into position toward {a}, no pause",
                f"immediately begins {a}, brisk continuous motion, no slow motion",
                f"presses {a} without stopping, fast decisive movement",
            ]
            out = [templates[min(i, len(templates) - 1)] for i in range(num_clips - 1)]
            out.append(f"explosively completes {a} at real-time speed, no slow motion, no freeze")
        elif n <= num_clips:
            out = [""] * num_clips
            out[last] = fast(parts[-1], "explosively")
            head = parts[:-1]
            for i, ph in enumerate(head):
                s = min(i, num_clips - 2)
                out[s] = (out[s] + ", then " + fast(ph)).strip(", ") if out[s] else fast(ph)
            for i in range(num_clips - 1):
                if not out[i]:
                    key = parts[min(i, n - 1)]
                    out[i] = f"rapidly moves toward {key}, brisk continuous motion, no slow motion"
        else:
            k, m = divmod(n - 1, num_clips - 1)
            idx, chunks = 0, []
            for i in range(num_clips - 1):
                end = idx + k + (1 if i < m else 0)
                seg = parts[idx:end]
                idx = end
                chunks.append(", then ".join(fast(p) for p in seg) if seg else "")
            out = chunks + [fast(parts[-1], "explosively")]
        return out[:num_clips]

    def _load(self, name):
        return comfy.utils.load_torch_file(folder_paths.get_full_path("loras", name))

    @staticmethod
    def _parse_weights(text):
        try:
            nums = [float(x) for x in re.split(r'[,\s]+', text.strip()) if x != ""]
        except (ValueError, AttributeError):
            nums = []
        if not nums:
            return 1.0, 1.0, 1.0
        if len(nums) == 1:
            return nums[0], nums[0], nums[0]
        if len(nums) == 2:
            return nums[0], nums[1], nums[1]
        return nums[0], nums[1], nums[2]

    def _patch_clip(self, mb, mh, ml, slots, cache, inject):
        cb, ch, cl = mb, mh, ml
        used = []
        for (name, w_str) in slots:
            if name == "None":
                continue
            w_b, w_h, w_l = self._parse_weights(w_str)
            if w_b == 0.0 and w_h == 0.0 and w_l == 0.0:
                continue
            data = self._load(name)
            if w_b != 0.0:
                cb = apply_model_lora(cb, data, w_b)
            if w_h != 0.0:
                ch = apply_model_lora(ch, data, w_h)
            if w_l != 0.0:
                cl = apply_model_lora(cl, data, w_l)
            used.append(name)
        trig = []
        trig_dbg = []
        for n in dict.fromkeys(used):
            found = get_triggers(n, cache)
            trig.extend(found)
            trig_dbg.append(f"{os.path.basename(n)}:{len(found)}")
        self._last_trig_dbg = trig_dbg
        trig_str = ", ".join(dict.fromkeys(t for t in trig if t))
        prompt_trig = trig_str if inject else ""
        return cb, ch, cl, prompt_trig, trig_str

    def direct(self, image, model_base, model_high, model_low, essential_actions, num_clips,
               character_name, artistic_vibe, pacing, detail_level, inject_triggers,
               use_vision, openrouter_api_key, vlm_preset, vlm_model, llm_preset, llm_model,
               creativity, seed, **kw):

        n_clips = int(str(num_clips).split()[0])
        vlm = resolve_slug(VLM_PRESETS, vlm_preset, vlm_model, "qwen/qwen2.5-vl-72b-instruct")
        llm = resolve_slug(LLM_PRESETS, llm_preset, llm_model, "deepseek/deepseek-v3.2")
        user_name = "" if character_name.strip().upper() in ["AUTO", ""] else character_name.strip()
        key = resolve_api_key(openrouter_api_key)

        vision_caption = str(kw.get("vision_caption", "") or "").strip()
        anchors, dbg_v = "", ""
        if vision_caption:
            anchors = vision_caption
            dbg_v = "vision ok (local caption node)"
            print(f"👁 [Director] 로컬 캡션 입력 사용 (JoyCaption 등):\n{anchors[:300]}")
        elif not use_vision:
            dbg_v = "vision OFF (토글 확인)"
            print("ℹ️ [Director] use_vision=OFF - 비전 건너뜀")
        elif not key:
            dbg_v = "no api key"
            print("❌ [Director] openrouter_api_key 비어있음 (위젯도, OPENROUTER_API_KEY 환경변수도 없음) - 비전/LLM 불가")
        elif image is None:
            dbg_v = "no image"
            print("❌ [Director] image 입력 없음 - IMAGE 연결 확인")
        else:
            data_uri = tensor_to_base64(image)
            if data_uri:
                anchors, err = self._vision_analyze(data_uri, vlm, key, min(creativity, 0.5))
                dbg_v = err or "vision ok"
            else:
                dbg_v = "image encode failed / PIL missing"
        if not vision_caption and use_vision and key and image is not None and not anchors:
            print(f"⚠️ [Director] 비전 앵커가 비었음 -> 이름/씬 근거 없음. 원인: {dbg_v}")

        if key:
            parsed, dbg_p = self._plan_shots(anchors, essential_actions, character_name,
                                             artistic_vibe, pacing, self._detail_hint(detail_level),
                                             llm, key, creativity, seed, num_clips=n_clips)
            if parsed:
                clips = [str(parsed[f"clip_{i}"]).strip() for i in range(1, n_clips + 1)]
                master = str(parsed.get("master_scene", "")).strip()
                cand = str(parsed.get("character", "")).strip()
                if not cand or cand.upper() == "AUTO":
                    cand = self._fallback_name(anchors)
                final_name = user_name or cand
            else:
                clips = self._local_split(essential_actions, num_clips=n_clips)
                master = anchors.replace("\n", " ")[:160]
                final_name = user_name or self._fallback_name(anchors)
        else:
            parsed, dbg_p = None, "no api key"
            clips = self._local_split(essential_actions, num_clips=n_clips)
            master = anchors.replace("\n", " ")[:160]
            final_name = user_name

        if len(clips) < 4:
            clips = clips + [clips[-1]] * (4 - len(clips))

        cache = _load_cache()
        manual_list = [t.strip() for t in kw.get("manual_triggers", "").split(',') if t.strip()]
        outs = []
        all_trigs_collected = []
        for c in range(1, 5):
            if c > n_clips:
                cb = ch = cl = ExecutionBlocker(None)
                prompt = self._build_final(clips[c - 1], master, final_name, "")
                outs.extend([cb, ch, cl, prompt])
                print(f"  [clip{c}] SKIPPED (num_clips={n_clips}) - ExecutionBlocker 반환")
                continue
            slots = [(kw.get(f"c{c}_lora_{n}", "None"),
                      kw.get(f"c{c}_w_{n}", "1.0, 1.0, 1.0")) for n in range(1, 5)]
            cb, ch, cl, prompt_trig, always_trig = self._patch_clip(
                model_base, model_high, model_low, slots, cache, inject_triggers)
            all_trig = ", ".join(dict.fromkeys(
                ([prompt_trig] if prompt_trig else []) + manual_list)) \
                if (prompt_trig or manual_list) else ""
            if always_trig:
                all_trigs_collected.extend(t.strip() for t in always_trig.split(',') if t.strip())
            prompt = self._build_final(clips[c - 1], master, final_name, all_trig)
            outs.extend([cb, ch, cl, prompt])
            print(f"  [clip{c}] loras/triggers: {getattr(self, '_last_trig_dbg', [])} "
                  f"inject={inject_triggers}")

        triggers_out = ", ".join(dict.fromkeys(
            [t for t in all_trigs_collected if t] + manual_list))

        debug = f"vlm={dbg_v} | plan={dbg_p} | char={final_name} | clips={n_clips} ({n_clips*5}s)"
        print(f"\n🎬 [Dolphin Vision Director] {debug}\nANCHORS: {anchors[:120]}")
        for i in range(4):
            print(f"  CLIP {i+1}: {outs[i*4+3][:90]}")

        is_final = tuple(c == n_clips for c in range(1, 5))

        return tuple(outs) + (triggers_out, anchors, debug) + is_final


NODE_CLASS_MAPPINGS = {"DolphinVisionDirector": DolphinVisionDirector}
NODE_DISPLAY_NAME_MAPPINGS = {"DolphinVisionDirector": "🎬👁 Dolphin Vision Director (Image→4Clips)"}