From ccfa97d03bb7406cb032d1f5f8e455ac695e2bf8 Mon Sep 17 00:00:00 2001 From: Hawk Lee Date: Mon, 29 Dec 2025 17:52:12 +0800 Subject: [PATCH] Fix 7B path (vibevoice/VibeVoice-7B), add max_length_times param to prevent early cutoff --- README.md | 7 ++++--- aiia_vibevoice_nodes.py | 6 ++++-- 2 files changed, 8 insertions(+), 5 deletions(-) diff --git a/README.md b/README.md index aa266ed..6f60391 100644 --- a/README.md +++ b/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）。 diff --git a/aiia_vibevoice_nodes.py b/aiia_vibevoice_nodes.py index 9960829..4b313c7 100644 --- a/aiia_vibevoice_nodes.py +++ b/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)