Fix 7B path (vibevoice/VibeVoice-7B), add max_length_times param to prevent early cutoff

README.md

@@ -432,16 +432,17 @@ git clone https://github.com/havvk/ComfyUI_AIIA.git
 - **当前状态**: ✅ 可用（已通过测试）
 - **支持语言**: **英文 (en) 和 中文 (zh)**（官方仅在这两种语言数据集上训练）
 - **可选模型**:
-  - `microsoft/VibeVoice-1.5B`: 轻量版，适合大多数场景（~3GB 显存）
-  - `microsoft/VibeVoice-7B`: 高质量版，效果更好但需要更多显存（~14GB）
+  - `microsoft/VibeVoice-1.5B`: 轻量版，64K 上下文（~3GB 显存）
+  - `vibevoice/VibeVoice-7B`: 高质量版，32K 上下文（~14GB 显存）
 - **特点**:
   - **即时启动**: 无需预热或编译，首次运行即可使用（CosyVoice 首次需 ~1 分钟编译）。
   - **语言自动识别**: 模型会自动识别中英文文本。
   - **零样本音色克隆**: 输入 `reference_audio` 即可克隆声音。
 - **节点参数**:
-  - `cfg_scale` (默认: 3.0): CFG 引导强度。值越高，语音越忠实于文本内容。
+  - `cfg_scale` (默认: 3.0): CFG 引导强度。值越高，语音越忠实于文本内容，**高 CFG 音质远超 CosyVoice**。
   - `ddpm_steps` (默认: 50): 扩散推理步数。越高质量越好但越慢（推荐 30-100）。
   - `speed` (默认: 1.0): 播放速度。>1 = 更快, <1 = 更慢（后处理时间拉伸，保持音调）。
+  - `max_length_times` (默认: 5.0): 最大生成长度 = 输入长度 × 此值。如音频提前结束请增大此值。
 - **环境要求**:
   - **Flash Attention 2**: 强烈推荐安装（否则速度较慢）。
   - **Transformers**: `>= 4.51`（重要: 旧版本不支持该模型）。

aiia_vibevoice_nodes.py

@@ -10,7 +10,7 @@ class AIIA_VibeVoice_Loader:
     def INPUT_TYPES(cls):
         return {
             "required": {
-                "model_name": (["microsoft/VibeVoice-1.5B", "microsoft/VibeVoice-7B"],),
+                "model_name": (["microsoft/VibeVoice-1.5B", "vibevoice/VibeVoice-7B"],),
                 "precision": (["fp16", "bf16", "fp32"], {"default": "fp16"}),
             }
         }
@@ -398,6 +398,7 @@ class AIIA_VibeVoice_TTS:
                 "cfg_scale": ("FLOAT", {"default": 3.0, "min": 1.0, "max": 10.0, "step": 0.5, "tooltip": "CFG scale for speech generation. Higher = more faithful to text."}),
                 "ddpm_steps": ("INT", {"default": 50, "min": 10, "max": 100, "step": 10, "tooltip": "Diffusion steps. Higher = better quality but slower."}),
                 "speed": ("FLOAT", {"default": 1.0, "min": 0.5, "max": 2.0, "step": 0.1, "tooltip": "Playback speed. >1 = faster, <1 = slower (post-process time-stretch)."}),
+                "max_length_times": ("FLOAT", {"default": 5.0, "min": 2.0, "max": 20.0, "step": 1.0, "tooltip": "Max generation length = input_length × this. Increase if audio cuts off early."}),
             },
             "optional": {
                 "reference_audio": ("AUDIO",),
@@ -409,7 +410,7 @@ class AIIA_VibeVoice_TTS:
     FUNCTION = "generate"
     CATEGORY = "AIIA/VibeVoice"
 
-    def generate(self, vibevoice_model, text, cfg_scale, ddpm_steps, speed, reference_audio=None):
+    def generate(self, vibevoice_model, text, cfg_scale, ddpm_steps, speed, max_length_times, reference_audio=None):
         model = vibevoice_model["model"]
         tokenizer = vibevoice_model["tokenizer"]
         processor = vibevoice_model.get("processor")
@@ -484,6 +485,7 @@ class AIIA_VibeVoice_TTS:
                  "eos_token_id": tokenizer.eos_token_id, 
                  "pad_token_id": tokenizer.eos_token_id,
                  "cfg_scale": cfg_scale, # User-controlled CFG scale for speech generation
+                 "max_length_times": max_length_times, # Control max generation length (input_length × this)
              }
              
              # Set diffusion inference steps (crucial for quality/speed tradeoff)