feat: 新增首尾帧 K3 自研节点、代理参数重命名、Gemini 客户端优化

- 新增 K3VideoFirstLast 节点（首尾帧 K3 自研）：基于 K3Video 去掉分镜，新增可选尾帧输入，尾帧通过 metadata.image_tail 传递
- nano_banana_pro / batch_nano_banana_pro：将参数「代理加速」重命名为「代理端口（如7897）」
- gemini_client：新增 build_proxy_url、_estimate_body_size、_scale_images_to_fit 工具方法
- base_client：小幅优化

__init__.py

@@ -22,7 +22,7 @@ except Exception:
 import ssl
 
 from .nodes import NanoBananaPro, BatchNanoBananaPro, GoogleGemini, LoadFile, ImageStitchPro, SaveCleanImage, BatchCleanMetadata, VideoPreview, GoogleVeo, FluxImageEdit, UniversalLLMChat, KlingVideo, KlingFirstLastFrame, KlingMotionControlTest, AspectRatioPreset, MultiResPreview, BatchImagesO1key, Seedance, SeedanceMultiModal, StreamPreview, DoubaoImage, O1keyGPTImage, KVideo
-from .nodes import K3Video
+from .nodes import K3Video, K3VideoFirstLast
 
 # 报错弹框友好文案（不修改原节点代码，仅在外层统一处理）
 _MSG_TIMEOUT = "API 请求超时，请稍后重试或检查网络。"
@@ -83,6 +83,7 @@ NODE_CLASS_MAPPINGS = {
     "O1keyGPTImage": O1keyGPTImage,
     "KVideo": KVideo,
     "K3Video": K3Video,
+    "K3VideoFirstLast": K3VideoFirstLast,
 }
 
 NODE_DISPLAY_NAME_MAPPINGS = {
@@ -110,6 +111,7 @@ NODE_DISPLAY_NAME_MAPPINGS = {
     "O1keyGPTImage": "o1key GPT Image",
     "KVideo": "K26 图生视频",
     "K3Video": "K3 图生视频 自研",
+    "K3VideoFirstLast": "首尾帧 K3 自研",
 }
 
 WEB_DIRECTORY = "./web"

clients/base_client.py

@@ -43,6 +43,7 @@ class BaseAPIClient(ABC):
         self.base_url = base_url
         self.api_key = api_key
         self.max_request_size = max_request_size
+        self.proxy_url: Optional[str] = None  # 由节点在调用前注入，如 "http://127.0.0.1:7897"
     
     @abstractmethod
     def get_endpoint(self, **kwargs) -> str:
@@ -200,7 +201,7 @@ class BaseAPIClient(ABC):
 
         async def _do_request():
             connect_start = time.time()
-            async with session.post(url, json=request_body, headers=headers, timeout=_aiohttp_timeout) as response:
+            async with session.post(url, json=request_body, headers=headers, timeout=_aiohttp_timeout, proxy=self.proxy_url) as response:
                 connect_time = time.time() - connect_start
 
                 if response.status != 200:

clients/gemini_client.py

@@ -38,6 +38,23 @@ class GeminiAPIClient(BaseAPIClient):
             max_request_size=100 * 1024 * 1024
         )
     
+    @staticmethod
+    def build_proxy_url(port: str) -> Optional[str]:
+        """
+        将端口号字符串转为 aiohttp 可用的 HTTP 代理 URL。
+        兼容 Windows / Mac，支持 v2rayN (10808) 和 Clash Verge (7897)。
+
+        Args:
+            port: 用户填写的端口号，如 "7897" 或 "10808"；空字符串返回 None
+
+        Returns:
+            代理 URL，如 "http://127.0.0.1:7897"，或 None（不使用代理）
+        """
+        port = (port or "").strip()
+        if not port or not port.isdigit():
+            return None
+        return f"http://127.0.0.1:{port}"
+
     def get_endpoint(self, model: str = "", resolution: str = "2K", **kwargs) -> str:
         """
         根据模型和分辨率获取 API 端点
@@ -124,6 +141,37 @@ class GeminiAPIClient(BaseAPIClient):
             return "服务无法在截止期限内完成处理。可能原因是：您的提示词过大，无法及时处理。"
         return None
     
+    @staticmethod
+    def _estimate_body_size(parts: list, extra_body: dict) -> int:
+        """
+        快速估算请求体 JSON 序列化后的字节数。
+        extra_body 为除 contents[0].parts 以外的其余字段。
+        """
+        import json
+        body = {
+            "contents": [{"role": "user", "parts": parts}],
+            **extra_body
+        }
+        return len(json.dumps(body).encode("utf-8"))
+
+    @staticmethod
+    def _scale_images_to_fit(
+        images: List[Image.Image],
+        scale: float
+    ) -> List[Image.Image]:
+        """
+        将所有图片按统一比例等比缩放（Lanczos，不降质量）。
+        scale < 1.0 时缩小，>= 1.0 时原样返回。
+        """
+        if scale >= 1.0:
+            return images
+        result = []
+        for img in images:
+            new_w = max(1, int(img.width * scale))
+            new_h = max(1, int(img.height * scale))
+            result.append(img.resize((new_w, new_h), Image.Resampling.LANCZOS))
+        return result
+
     def build_request_body(
         self,
         prompt: str = "",
@@ -148,6 +196,8 @@ class GeminiAPIClient(BaseAPIClient):
         Returns:
             请求体字典
         """
+        _MAX_BODY_BYTES = 20 * 1024 * 1024  # 20 MB
+
         parts = []
 
         # 添加文本部分
@@ -155,14 +205,59 @@ class GeminiAPIClient(BaseAPIClient):
 
         # 添加图像部分（如果有）
         if images:
-            for img in images:
-                img_base64 = encode_image_to_base64(img)
-                parts.append({
-                    "inline_data": {
-                        "mime_type": "image/png",
-                        "data": img_base64
+            working_images = list(images)
+
+            # 编码一次，估算大小，超限则迭代缩放
+            for _attempt in range(10):
+                img_parts = []
+                for img in working_images:
+                    img_base64 = encode_image_to_base64(img)
+                    img_parts.append({
+                        "inline_data": {
+                            "mime_type": "image/png",
+                            "data": img_base64
+                        }
+                    })
+
+                # 估算完整 body 大小（不含工具字段，工具字段很小可忽略）
+                estimated = self._estimate_body_size(
+                    parts + img_parts,
+                    {
+                        "generationConfig": {
+                            "responseModalities": ["TEXT", "IMAGE"],
+                            "imageConfig": {
+                                "aspectRatio": aspect_ratio,
+                                "imageSize": resolution
+                            }
+                        }
                     }
-                })
+                )
+
+                if estimated <= _MAX_BODY_BYTES:
+                    parts.extend(img_parts)
+                    if _attempt > 0:
+                        orig_sizes = ", ".join(
+                            f"{img.width}×{img.height}" for img in images
+                        )
+                        new_sizes = ", ".join(
+                            f"{img.width}×{img.height}" for img in working_images
+                        )
+                        size_mb = estimated / (1024 * 1024)
+                        print(
+                            f"Nano Banana Pro: 输入图片已自动缩放以控制请求体积\n"
+                            f"  原始尺寸: {orig_sizes}\n"
+                            f"  缩放后:   {new_sizes}\n"
+                            f"  请求体积: {size_mb:.2f}MB（限制 20MB）"
+                        )
+                    break
+                else:
+                    # 按像素面积比推算需要的线性缩放系数，留 5% 余量
+                    ratio = (_MAX_BODY_BYTES * 0.95) / estimated
+                    scale = ratio ** 0.5  # 面积比 → 线性比
+                    working_images = self._scale_images_to_fit(working_images, scale)
+            else:
+                # 10 轮后仍超限，使用最后一次结果（极端情况兜底）
+                parts.extend(img_parts)
 
         # 构建请求体
         request_body = {

nodes/K3_video_firstlast.py

@@ -0,0 +1,307 @@
+"""
+首尾帧 K3 自研节点
+基于 K3 图生视频 自研，去掉分镜功能，新增尾帧可选输入。
+"""
+
+import asyncio
+import json
+import os
+import re
+
+import aiohttp
+
+from ..utils.config import get_api_key_or_raise, get_api_base_url
+from ..utils.image_utils import tensor_to_pil, encode_image_to_base64
+
+try:
+    from comfy_api.latest import InputImpl
+    import folder_paths
+    _FOLDER_PATHS_OK = True
+except Exception:
+    _FOLDER_PATHS_OK = False
+
+
+# ── 常量 ──────────────────────────────────────────────────────────────────────
+
+_MODEL_BASE = "kling-v3"
+_MODES      = ["标准", "专家", "4K"]
+_MODE_MAP   = {"标准": "std", "专家": "pro", "4K": "4k"}
+
+_ENDPOINT_CREATE = "/v1/video/generations"
+_ENDPOINT_STATUS = "/v1/video/generations/{task_id}"
+
+_POLL_INIT = 3
+_POLL_MAX  = 15
+
+
+# ── 工具函数 ───────────────────────────────────────────────────────────────────
+
+def _get_video_dir() -> str:
+    if _FOLDER_PATHS_OK:
+        base = folder_paths.get_output_directory()
+    else:
+        plugin = os.path.dirname(os.path.dirname(os.path.abspath(__file__)))
+        base = os.path.join(os.path.dirname(os.path.dirname(plugin)), "output")
+    d = os.path.join(base, "video")
+    os.makedirs(d, exist_ok=True)
+    return d
+
+
+def _next_counter(directory: str, prefix: str) -> int:
+    pattern = re.compile(rf"^{re.escape(prefix)}_(\d+)")
+    max_n = 0
+    if os.path.exists(directory):
+        for f in os.listdir(directory):
+            m = pattern.match(f)
+            if m:
+                max_n = max(max_n, int(m.group(1)))
+    return max_n + 1
+
+
+def _prepare_image_base64(tensor) -> str:
+    """转换并校验图片，不符合约束时自动等比缩放后返回 base64。"""
+    import io
+    import base64
+
+    pil_list = tensor_to_pil(tensor)
+    img = pil_list[0].convert("RGB")
+    w, h = img.size
+
+    # 1. 宽高比校验
+    ratio = w / h
+    if ratio < 1 / 2.5 or ratio > 2.5:
+        raise RuntimeError(
+            f"图片宽高比 {w}:{h}（{ratio:.2f}）超出允许范围 1:2.5 ~ 2.5:1，请裁剪后重试。"
+        )
+
+    # 2. 最小尺寸：任意边 < 300px 时等比放大
+    if w < 300 or h < 300:
+        scale = max(300 / w, 300 / h)
+        img = img.resize((int(w * scale), int(h * scale)), resample=1)
+
+    # 3. 文件大小：循环等比缩小直到 ≤ 10MB
+    MAX_BYTES = 10 * 1024 * 1024
+    for _ in range(20):
+        buf = io.BytesIO()
+        img.save(buf, format="PNG")
+        if buf.tell() <= MAX_BYTES:
+            break
+        scale = (MAX_BYTES / buf.tell()) ** 0.5 * 0.95
+        new_w = int(img.width * scale)
+        new_h = int(img.height * scale)
+        if new_w < 300 or new_h < 300:
+            raise RuntimeError(
+                f"图片压缩至 10MB 以内后尺寸（{new_w}x{new_h}）低于最小限制 300px，无法同时满足两项约束。"
+            )
+        img = img.resize((new_w, new_h), resample=1)
+    else:
+        raise RuntimeError("图片经过 20 次缩放仍超过 10MB，请检查原始图片。")
+
+    buf.seek(0)
+    return base64.b64encode(buf.read()).decode("utf-8")
+
+
+# ── 节点 ──────────────────────────────────────────────────────────────────────
+
+class K3VideoFirstLast:
+    """首尾帧 K3 自研"""
+
+    @classmethod
+    def INPUT_TYPES(cls):
+        return {
+            "required": {
+                "起始帧":    ("IMAGE",),
+                "提示词":    ("STRING", {"multiline": True, "default": ""}),
+                "负向提示词": ("STRING", {"multiline": True, "default": ""}),
+                "时长":      ([5, 10, 15], {"default": 5}),
+                "生成音频":  (["关闭", "打开"], {"default": "关闭"}),
+                "模式":      (_MODES, {"default": "标准"}),
+                "seed": ("INT", {
+                    "default": 0, "min": 0, "max": 2147483647,
+                    "tooltip": "seed 仅控制节点是否重新运行，结果本身不可复现。",
+                }),
+            },
+            "optional": {
+                "尾帧": ("IMAGE", {"tooltip": "可选。传入后将作为视频尾帧参考。"}),
+            },
+        }
+
+    RETURN_TYPES  = ("VIDEO",)
+    RETURN_NAMES  = ("视频",)
+    FUNCTION      = "generate"
+    CATEGORY      = "comfyui_o1key/KVideo"
+
+    async def generate(self, 起始帧, 提示词, 负向提示词, 时长, 生成音频, 模式, seed, 尾帧=None):
+        api_key  = get_api_key_or_raise()
+        base_url = get_api_base_url()
+        headers  = {
+            "Authorization": f"Bearer {api_key}",
+            "Content-Type":  "application/json",
+        }
+
+        voice      = "voice" if 生成音频 == "打开" else "novoice"
+        mode_api   = _MODE_MAP[模式]
+        if mode_api == "4k":
+            model_name = f"{_MODEL_BASE}-4k-{时长}s"
+        else:
+            model_name = f"{_MODEL_BASE}-{mode_api}-{时长}s-{voice}"
+
+        if not 提示词.strip():
+            raise RuntimeError("提示词不能为空。")
+
+        # ── 构建请求体 ────────────────────────────────────────────────
+        body: dict = {
+            "model":    model_name,
+            "prompt":   提示词.strip(),
+            "mode":     mode_api,
+            "duration": 时长,
+            "image":    _prepare_image_base64(起始帧),
+        }
+
+        if 负向提示词.strip():
+            body["negative_prompt"] = 负向提示词.strip()
+
+        # metadata：尾帧 + 音频
+        metadata: dict = {}
+        if 尾帧 is not None:
+            metadata["image_tail"] = _prepare_image_base64(尾帧)
+        if 生成音频 == "打开":
+            metadata["sound"] = "on"
+        if metadata:
+            body["metadata"] = metadata
+
+        # generate_audio 字段（非 metadata 路径）
+        if 生成音频 == "打开" and not metadata.get("sound"):
+            body["generate_audio"] = True
+
+        # ── 进度条 ────────────────────────────────────────────────────
+        try:
+            from comfy.utils import ProgressBar
+            pbar = ProgressBar(100)
+        except Exception:
+            pbar = None
+
+        def _stage(s: str):
+            if s == "submitting":
+                print("[K3 首尾帧] 提交中...")
+                if pbar: pbar.update_absolute(0, 100)
+            elif s.startswith("submitted:"):
+                print(f"[K3 首尾帧] 任务已提交 → {s.split(':', 1)[1]}")
+                if pbar: pbar.update_absolute(5, 100)
+            elif s == "downloading":
+                print("[K3 首尾帧] 下载视频...")
+                if pbar: pbar.update_absolute(99, 100)
+            elif s == "done":
+                print("[K3 首尾帧] 完成")
+                if pbar: pbar.update_absolute(100, 100)
+
+        def _progress(pct: int):
+            if pbar: pbar.update_absolute(5 + int(pct * 0.94), 100)
+
+        # ── 保存路径 ──────────────────────────────────────────────────
+        video_dir = _get_video_dir()
+        counter   = _next_counter(video_dir, "k3fl")
+        save_path = os.path.join(video_dir, f"k3fl_{counter:05d}.mp4")
+
+        connector = aiohttp.TCPConnector(ssl=False, force_close=True)
+        async with aiohttp.ClientSession(connector=connector) as session:
+
+            # 1. 提交
+            _stage("submitting")
+            create_url = f"{base_url}{_ENDPOINT_CREATE}"
+            async with session.post(create_url, json=body, headers=headers) as resp:
+                text = await resp.text()
+                if resp.status != 200:
+                    try:
+                        err = json.loads(text)
+                        msg = err.get("error", {}).get("message") or err.get("message") or text
+                    except Exception:
+                        msg = text
+                    raise RuntimeError(f"K3 首尾帧提交失败 ({resp.status}): {msg}")
+                create_resp = json.loads(text)
+
+            task_id = (
+                create_resp.get("task_id")
+                or create_resp.get("id")
+                or create_resp.get("data", {}).get("task_id")
+            )
+            if not task_id:
+                raise RuntimeError(f"API 未返回任务 ID，响应：{create_resp}")
+            _stage(f"submitted:{task_id}")
+
+            # 2. 轮询
+            status_url = f"{base_url}{_ENDPOINT_STATUS.format(task_id=task_id)}"
+            interval   = _POLL_INIT
+            video_url  = None
+
+            while True:
+                async with session.get(status_url, headers=headers) as resp:
+                    text = await resp.text()
+                    if resp.status != 200:
+                        try:
+                            err = json.loads(text)
+                            msg = err.get("error", {}).get("message") or err.get("message") or text
+                        except Exception:
+                            msg = text
+                        raise RuntimeError(f"状态查询失败 ({resp.status}): {msg}")
+                    sr = json.loads(text)
+
+                data   = sr.get("data", sr)
+                status = (data.get("status") or sr.get("status") or "").lower()
+
+                pct_raw = data.get("progress", 0)
+                try:
+                    pct = int(str(pct_raw).rstrip("%").strip())
+                except (ValueError, AttributeError):
+                    pct = 0
+                print(f"[K3 首尾帧] 生成中 {pct}%")
+                _progress(pct)
+
+                if status in ("success", "completed", "done", "finished", "succeed"):
+                    video_url = (
+                        data.get("video_url")
+                        or data.get("result_url")
+                        or data.get("url")
+                        or (data.get("result", {}) or {}).get("url")
+                        or sr.get("video_url")
+                        or sr.get("url")
+                    )
+                    break
+                if status in ("failed", "fail"):
+                    err_info = data.get("error") or sr.get("error") or {}
+                    err_msg  = (err_info.get("message", "未知错误")
+                                if isinstance(err_info, dict) else str(err_info))
+                    raise RuntimeError(f"K3 首尾帧生成失败：{err_msg}")
+
+                await asyncio.sleep(interval)
+                interval = min(interval * 1.5, _POLL_MAX)
+
+            if not video_url:
+                raise RuntimeError(f"API 未返回视频 URL，响应：{sr}")
+
+            # 3. 下载
+            _stage("downloading")
+            async with session.get(video_url, allow_redirects=True) as resp:
+                if resp.status != 200:
+                    raise RuntimeError(f"视频下载失败 ({resp.status})")
+                os.makedirs(os.path.dirname(save_path), exist_ok=True)
+                with open(save_path, "wb") as f:
+                    async for chunk in resp.content.iter_chunked(8192):
+                        f.write(chunk)
+
+        _stage("done")
+
+        if _FOLDER_PATHS_OK:
+            return (InputImpl.VideoFromFile(save_path),)
+        return (save_path,)
+
+
+# ── 节点注册 ──────────────────────────────────────────────────────────────────
+
+NODE_CLASS_MAPPINGS = {
+    "K3VideoFirstLast": K3VideoFirstLast,
+}
+
+NODE_DISPLAY_NAME_MAPPINGS = {
+    "K3VideoFirstLast": "首尾帧 K3 自研",
+}

nodes/__init__.py

@@ -22,5 +22,6 @@ from .doubao_image import DoubaoImage
 from .gpt_image import O1keyGPTImage
 from .K_video import KVideo
 from .K3_video import K3Video
+from .K3_video_firstlast import K3VideoFirstLast
 
-__all__ = ['NanoBananaPro', 'BatchNanoBananaPro', 'GoogleGemini', 'LoadFile', 'ImageStitchPro', 'SaveCleanImage', 'BatchCleanMetadata', 'VideoPreview', 'KlingVideo', 'KlingFirstLastFrame', 'KlingMotionControlTest', 'AspectRatioPreset', 'GoogleVeo', 'FluxImageEdit', 'UniversalLLMChat', 'MultiResPreview', 'BatchImagesO1key', 'Seedance', 'SeedanceMultiModal', 'StreamPreview', 'DoubaoImage', 'O1keyGPTImage', 'KVideo', 'K3Video']
+__all__ = ['NanoBananaPro', 'BatchNanoBananaPro', 'GoogleGemini', 'LoadFile', 'ImageStitchPro', 'SaveCleanImage', 'BatchCleanMetadata', 'VideoPreview', 'KlingVideo', 'KlingFirstLastFrame', 'KlingMotionControlTest', 'AspectRatioPreset', 'GoogleVeo', 'FluxImageEdit', 'UniversalLLMChat', 'MultiResPreview', 'BatchImagesO1key', 'Seedance', 'SeedanceMultiModal', 'StreamPreview', 'DoubaoImage', 'O1keyGPTImage', 'KVideo', 'K3Video', 'K3VideoFirstLast']

nodes/batch_nano_banana_pro.py

@@ -214,6 +214,12 @@ class BatchNanoBananaPro:
         optional_inputs["图片配对模式"] = (cls.PAIRING_MODES, {
             "default": "不配对"
         })
+
+        optional_inputs["代理加速"] = ("STRING", {
+            "default": "",
+            "multiline": False,
+            "placeholder": "本地代理端口，如 7897（Clash Verge）或 10808（v2rayN），留空不使用"
+        })
         
         return {
             "required": {
@@ -742,6 +748,7 @@ class BatchNanoBananaPro:
         # 从 kwargs 提取搜索参数（界面显示为「关闭/打开」，转为 bool 供调用）
         enable_grounding: bool = (kwargs.pop("谷歌搜索（联网）", "关闭") == "打开")
         enable_image_search: bool = (kwargs.pop("图片搜索（联网）", "关闭") == "打开")
+        proxy_port: str = kwargs.pop("代理加速", "")
 
 
         try:
@@ -881,6 +888,11 @@ class BatchNanoBananaPro:
                     self.client = GeminiAPIClient()
                 except ValueError as e:
                     raise ValueError(f"初始化 API 客户端失败: {str(e)}")
+
+            # 注入代理设置（每次执行都刷新，支持用户中途修改端口）
+            self.client.proxy_url = GeminiAPIClient.build_proxy_url(proxy_port)
+            if self.client.proxy_url:
+                print(f"BatchNanoBananaPro: 已启用代理加速 → {self.client.proxy_url}")
             
             # 判断是否使用默认 output 目录
             original_save_path = kwargs.get('保存路径', '')

nodes/nano_banana_pro.py

@@ -149,6 +149,12 @@ class NanoBananaPro:
         optional_inputs = {}
         for i in range(1, 10):  # 1-9
             optional_inputs[f"参考图{i}"] = ("IMAGE",)
+
+        optional_inputs["代理端口（如7897）"] = ("STRING", {
+            "default": "",
+            "multiline": False,
+            "placeholder": "本地代理端口，如 7897（Clash Verge）或 10808（v2rayN），留空不使用"
+        })
         
         return {
             "required": {
@@ -464,6 +470,7 @@ class NanoBananaPro:
         # 从 kwargs 提取搜索参数（界面显示为「关闭/打开」，转为 bool 供调用）
         enable_grounding: bool = (kwargs.pop("谷歌搜索（联网）", "关闭") == "打开")
         enable_image_search: bool = (kwargs.pop("图片搜索（联网）", "关闭") == "打开")
+        proxy_port: str = kwargs.pop("代理端口（如7897）", "")
 
         # 创建 ComfyUI 原生进度条
         pbar = None
@@ -488,6 +495,11 @@ class NanoBananaPro:
                     self.client = GeminiAPIClient()
                 except ValueError as e:
                     raise ValueError(f"初始化失败: {str(e)}")
+
+            # 注入代理设置（每次执行都刷新，支持用户中途修改端口）
+            self.client.proxy_url = GeminiAPIClient.build_proxy_url(proxy_port)
+            if self.client.proxy_url:
+                print(f"Nano Banana Pro: 已启用代理加速 → {self.client.proxy_url}")
             
             # 校验分辨率与模型的兼容性
             supported_resolutions = get_model_supported_resolutions(模型)