Merge pull request #348 from shadowcz007/whisper-sensevoice

Whisper sensevoice

README.md

@@ -10,6 +10,12 @@ For business cooperation, please contact email 389570357@qq.com
 
 ##### `最新`：
 
+- 新增 SenseVoice
+
+- [新增JS-SDK，方便直接在前端项目中使用comfyui](https://github.com/shadowcz007/comfyui-js-sdk)
+
+- 新增API调用图像生成节点 TextToImage Siliconflow，可以直接调用Siliconflow提供的flux生成图像
+
 - [增加 Her 的DEMO页面，和数字人对话](https://github.com/shadowcz007/ComfyUI-Backend-MixlabNodes/blob/main/workflow/her_demo_workflow.json)
 
 - 右键菜单支持 text-to-text，方便对 prompt 词补全，支持云LLM或者是本地LLM。

__init__.py

@@ -1006,8 +1006,8 @@ from .nodes.ImageNode import DepthViewer_,ImageBatchToList_,ImageListToBatch_,Co
 # from .nodes.Vae import VAELoader,VAEDecode
 from .nodes.ScreenShareNode import ScreenShareNode,FloatingVideo
 
-from .nodes.Audio import AudioPlayNode,SpeechRecognition,SpeechSynthesis
-from .nodes.Utils import KeyInput,IncrementingListNode,ListSplit,CreateLoraNames,CreateSampler_names,CreateCkptNames,CreateSeedNode,TESTNODE_,TESTNODE_TOKEN,AppInfo,IntNumber,FloatSlider,TextInput,ColorInput,FontInput,TextToNumber,DynamicDelayProcessor,LimitNumber,SwitchByIndex,MultiplicationNode
+from .nodes.Audio import AudioPlayNode,SpeechRecognition,SpeechSynthesis,AnalyzeAudioNone
+from .nodes.Utils import CreateJsonNode,KeyInput,IncrementingListNode,ListSplit,CreateLoraNames,CreateSampler_names,CreateCkptNames,CreateSeedNode,TESTNODE_,TESTNODE_TOKEN,AppInfo,IntNumber,FloatSlider,TextInput,ColorInput,FontInput,TextToNumber,DynamicDelayProcessor,LimitNumber,SwitchByIndex,MultiplicationNode
 from .nodes.Mask import PreviewMask_,MaskListReplace,MaskListMerge,OutlineMask,FeatheredMask
 
 from .nodes.Style import ApplyVisualStylePrompting,StyleAlignedReferenceSampler,StyleAlignedBatchAlign,StyleAlignedSampleReferenceLatents
@@ -1045,6 +1045,7 @@ NODE_CLASS_MAPPINGS = {
     "SaveImageToLocal":SaveImageToLocal,
     "SaveImageAndMetadata_":SaveImageAndMetadata,
     "ComparingTwoFrames_":ComparingTwoFrames,
+    "CreateJsonNode":CreateJsonNode,
 
     # Image
     "MirroredImage":MirroredImage,
@@ -1102,6 +1103,7 @@ NODE_CLASS_MAPPINGS = {
     "SpeechRecognition":SpeechRecognition,
     "SpeechSynthesis":SpeechSynthesis,
     "AudioPlay":AudioPlayNode,
+    "AnalyzeAudio":AnalyzeAudioNone,
 
     # Text
     "TextToNumber":TextToNumber,
@@ -1219,12 +1221,14 @@ NODE_DISPLAY_NAME_MAPPINGS = {
     "SpeechSynthesis":"SpeechSynthesis ♾️Mixlab",
     "SpeechRecognition":"SpeechRecognition ♾️Mixlab",
     "AudioPlay":"Preview Audio ♾️Mixlab",
+    "AnalyzeAudio":"Analyze Audio ♾️Mixlab",
 
     # Utils
     "DynamicDelayProcessor":"DynamicDelayByText ♾️Mixlab",
     "MultiplicationNode":"Math Operation ♾️Mixlab",
     "ListSplit_":"Split List ♾️Mixlab",
     "SwitchByIndex":"List Switch By Index ♾️Mixlab",
+    "CreateJsonNode":"Create Json",
 
     # "GamePal":"GamePal ♾️Mixlab",
     # Experiment
@@ -1427,6 +1431,24 @@ try:
 except Exception as e:
     logging.info('FishSpeech.available False' )
 
+try:
+    from .nodes.SenseVoice import SenseVoiceNode
+    logging.info('SenseVoice.available')
+    NODE_CLASS_MAPPINGS['SenseVoiceNode']=SenseVoiceNode
+    NODE_DISPLAY_NAME_MAPPINGS["SenseVoiceNode"]= "Sense Voice ♾️Mixlab"
 
+except Exception as e:
+    logging.info('SenseVoice.available False' )  
+
+try:
+    from .nodes.Whisper import LoadWhisperModel,WhisperTranscribe
+    logging.info('Whisper.available')
+    NODE_CLASS_MAPPINGS['LoadWhisperModel_']=LoadWhisperModel
+    NODE_CLASS_MAPPINGS['WhisperTranscribe_']=WhisperTranscribe
+    NODE_DISPLAY_NAME_MAPPINGS["LoadWhisperModel_"]= "Load Whisper Model ♾️Mixlab"
+    NODE_DISPLAY_NAME_MAPPINGS["WhisperTranscribe_"]= "Whisper Transcribe ♾️Mixlab"
+
+except Exception as e:
+    logging.info('Whisper.available False' )  
 
 logging.info('\033[93m -------------- \033[0m')

nodes/Audio.py

@@ -3,6 +3,101 @@ import os
 import folder_paths
 import torchaudio
 
+class AnyType(str):
+  """A special class that is always equal in not equal comparisons. Credit to pythongosssss"""
+
+  def __ne__(self, __value: object) -> bool:
+    return False
+
+any_type = AnyType("*")
+
+
+def analyze_audio_data(audio_data):
+    total_duration = 0
+    total_gap_duration = 0
+    emotion_counts = {}
+    audio_types = set()
+    languages = set()
+
+    for i, entry in enumerate(audio_data):
+        # Calculate the duration of each audio segment
+        start_time = entry['start_time']
+        end_time = entry['end_time']
+        duration = end_time - start_time
+        total_duration += duration
+
+        # Count the emotions
+        if "emotion" in entry:
+            emotion = entry['emotion']
+            if emotion in emotion_counts:
+                emotion_counts[emotion] += 1
+            else:
+                emotion_counts[emotion] = 1
+
+        # Collect the audio types
+        if "audio_type" in entry:
+            audio_types.add(entry['audio_type'])
+
+        if "language" in entry:
+            languages.add(entry['language'])
+
+        # Calculate gap duration if not the last entry
+        if i < len(audio_data) - 1:
+            next_start_time = audio_data[i + 1]['start_time']
+            gap_duration = next_start_time - end_time
+            if gap_duration > 0:
+                total_gap_duration += gap_duration
+
+    # Get the most frequent emotion
+    if len(emotion_counts.keys())>0:
+        most_frequent_emotion = max(emotion_counts, key=emotion_counts.get)
+    else:
+        most_frequent_emotion=None
+
+    # Convert audio_types set to list for better readability
+    audio_types = list(audio_types)
+
+    languages=list(languages)
+
+    # Print the results
+    print(f"Total Effective Duration: {total_duration:.2f} seconds")
+    print(f"Total Gap Duration: {total_gap_duration:.2f} seconds")
+    print(f"Emotion Changes: {emotion_counts}")
+    print(f"Most Frequent Emotion: {most_frequent_emotion}")
+    print(f"Audio Types: {audio_types}")
+
+
+    return {
+        "total_duration": total_duration,
+        "total_gap_duration": total_gap_duration,
+        "emotion_changes": emotion_counts,
+        "most_frequent_emotion": most_frequent_emotion,
+        "audio_types": audio_types,
+        "languages":languages
+    }
+
+
+# 分析音频数据
+class AnalyzeAudioNone:
+    @classmethod
+    def INPUT_TYPES(s):
+        return {"required": { 
+                    "json":(any_type,),}, 
+                }
+    
+    RETURN_TYPES = (any_type,)
+    RETURN_NAMES = ("result",)
+
+    FUNCTION = "run"
+
+    CATEGORY = "♾️Mixlab/Audio"
+
+    def run(self,json):
+        result=analyze_audio_data(json)
+        return (result,)
+
+
+
 class SpeechRecognition:
     @classmethod
     def INPUT_TYPES(s):

nodes/ChatGPT.py

@@ -460,7 +460,8 @@ class SiliconflowFreeNode:
     @classmethod
     def INPUT_TYPES(cls):
         model_list= [ 
-            "Qwen/Qwen2-7B-Instruct",
+            "Qwen/Qwen2.5-7B-Instruct",
+            "Qwen/Qwen2-7B-Instruct", 
             "THUDM/glm-4-9b-chat",
             "01-ai/Yi-1.5-9B-Chat-16K",
             "meta-llama/Meta-Llama-3.1-8B-Instruct"
@@ -785,9 +786,12 @@ class JsonRepair:
     def INPUT_TYPES(s):
         return {
             "required": {
-                 "json_string":("STRING", {"forceInput": True,}), 
-                 "key":("STRING", {"multiline": False,"dynamicPrompts": False,"default": ""}),
-            }
+                "json_string":("STRING", {"forceInput": True,}), 
+                "key":("STRING", {"multiline": False,"dynamicPrompts": False,"default": ""}),
+            },
+            "optional":{
+                "json_string2":("STRING", {"forceInput": True,})
+            },
         }
 
     INPUT_IS_LIST = False
@@ -799,8 +803,11 @@ class JsonRepair:
 
     CATEGORY = "♾️Mixlab/GPT"
 
-    def run(self, json_string,key=""):
+    def run(self, json_string,key="",json_string2=None):
 
+        if not isinstance(json_string, str):
+            json_string=json.dumps(json_string)
+    
         json_string=extract_json_strings(json_string)
         # print(json_string)
         good_json_string = repair_json(json_string)
@@ -808,6 +815,20 @@ class JsonRepair:
         # 将 JSON 字符串解析为 Python 对象
         data = json.loads(good_json_string)
 
+        if json_string2!=None:
+            if not isinstance(json_string2, str):
+                json_string2=json.dumps(json_string2)
+    
+            json_string2=extract_json_strings(json_string2)
+            # print(json_string)
+            good_json_string2 = repair_json(json_string2)
+
+            # 将 JSON 字符串解析为 Python 对象
+            data2 = json.loads(good_json_string2)
+
+            data={**data, **data2}
+
+
         v=""
         if key!="" and (key in data):
             v=data[key]

nodes/ImageNode.py

@@ -13,12 +13,26 @@ import json,io
 import comfy.utils
 from comfy.cli_args import args
 import cv2
-import string
+import string,re
 import math,glob
 from .Watcher import FolderWatcher
 
 from itertools import product
 
+
+# 文件名排序
+def sort_by_filename(items):
+    def extract_parts(filename):
+        # 使用正则表达式将文件名拆分为数字和非数字部分
+        parts = re.split(r'(\d+)', filename)
+        # 将数字部分转换为整数以便正确排序，同时保留非数字部分
+        parts = [int(part) if part.isdigit() else part for part in parts]
+        return parts
+
+    # 按照 file_name 的拆分部分进行排序
+    sorted_items = sorted(items, key=lambda x: extract_parts(x['file_name']))
+    return sorted_items
+
 # 将PIL图片转换为OpenCV格式
 def pil_to_opencv(image):
     open_cv_image = cv2.cvtColor(np.array(image), cv2.COLOR_RGB2BGR)
@@ -768,8 +782,7 @@ def areaToMask(x,y,w,h,image):
 #     return bg_image
 
 
-import cv2
-import numpy as np
+
 
 # ps的正片叠底
 # 可以基于https://www.cnblogs.com/jsxyhelu/p/16947810.html ，用gpt写python代码
@@ -1433,7 +1446,7 @@ class LoadImagesFromPath:
                             },
                 "optional":{
                     "white_bg": (["disable","enable"],),
-                    "newest_files": (["enable", "disable"],),
+                    "sort_by": (["file_name", "newest"],),#根据文件名来排序，还是按照最新创建时间
                     "index_variable":("INT", {
                         "default": 0, 
                         "min": -1, #Minimum value
@@ -1462,7 +1475,7 @@ class LoadImagesFromPath:
     watcher_folder=None
 
     # 运行的函数
-    def run(self,file_path,white_bg,newest_files,index_variable,watcher,result,prompt,seed=1):
+    def run(self,file_path,white_bg,sort_by,index_variable,watcher,result,prompt,seed=1):
         global watcher_folder
         # print('###监听:',watcher_folder,watcher,file_path,result)
 
@@ -1485,19 +1498,23 @@ class LoadImagesFromPath:
         # 当开启了监听，则取最新的，第一个文件
         if watcher=='enable':
             index_variable=0
-            newest_files='enable'
+            sort_by='newest'
 
         # 排序
-        sorted_files = sorted(images, key=lambda x: os.path.getmtime(x['file_path']), reverse=(newest_files=='enable'))
+        if sort_by=='newest':
+            sorted_files = sorted(images, key=lambda x: os.path.getmtime(x['file_path']), reverse=True)
+        elif sort_by=='file_name':
+            # 根据文件名排序
+            sorted_files = sort_by_filename(images)
 
         imgs=[]
         masks=[]
-        file_names=[]
+        file_paths=[]
 
         for im in sorted_files:
             imgs.append(im['image'])
             masks.append(im['mask'])
-            file_names.append(im['file_name'])
+            file_paths.append(im['file_path'])
         
         # print('index_variable',index_variable)
         
@@ -1505,13 +1522,13 @@ class LoadImagesFromPath:
             if index_variable!=-1:
                 imgs=[imgs[index_variable]] if index_variable < len(imgs) else None
                 masks=[masks[index_variable]] if index_variable < len(masks) else None
-                file_names=[file_names[index_variable]] if index_variable < len(file_names) else None
+                file_paths=[file_paths[index_variable]] if index_variable < len(file_paths) else None
         except Exception as e:
             print("发生了一个未知的错误：", str(e))
 
         # print('#prompt::::',prompt)
         # return  {"ui": {"seed": [1]}, "result":(imgs,masks,prompt,file_names,)}
-        return  (imgs,masks,prompt,file_names,)
+        return  (imgs,masks,prompt,file_paths,)
 
 
 # TODO 扩大选区的功能,重新输出mask
@@ -2873,14 +2890,14 @@ class ResizeImage:
                     "default": 512, 
                     "min": 1, #Minimum value
                     "max": 8192, #Maximum value
-                    "step": 8, #Slider's step
+                    "step": 1, #Slider's step
                     "display": "number" # Cosmetic only: display as "number" or "slider"
                 }),
                 "height": ("INT",{
                     "default": 512, 
                     "min": 1, #Minimum value
                     "max": 8192, #Maximum value
-                    "step": 8, #Slider's step
+                    "step": 1, #Slider's step
                     "display": "number" # Cosmetic only: display as "number" or "slider"
                 }),
                 "scale_option": (["width","height",'overall','center'],),

nodes/MiniCPMNode.py

@@ -1,4 +1,5 @@
 # Referenced some code：https://github.com/IuvenisSapiens/ComfyUI_MiniCPM-V-2_6-int4
+# https://github.com/CY-CHENYUE/ComfyUI-MiniCPM-Plus
 
 import os
 import torch
@@ -35,6 +36,7 @@ class MiniCPM_VQA_Simple:
                 "images": ("IMAGE",),
                 "text": ("STRING", {"default": "", "multiline": True}),   
                 "seed": ("INT", {"default": -1}),  # add seed parameter, default is -1
+                "extract_keywords":("BOOLEAN", {"default": False}),
                 "temperature": (
                     "FLOAT",
                     {
@@ -46,7 +48,9 @@ class MiniCPM_VQA_Simple:
             
         }
 
-    RETURN_TYPES = ("STRING",)
+    RETURN_TYPES = ("STRING","STRING",)
+    RETURN_NAMES = ("result","keywords",)
+
     FUNCTION = "inference"
     CATEGORY = "♾️Mixlab/Image"
 
@@ -55,6 +59,7 @@ class MiniCPM_VQA_Simple:
         images,
         text, 
         seed,  # add seed parameter, default is -1
+        extract_keywords,
         temperature,
         keep_model_loaded,
     ):
@@ -90,6 +95,7 @@ class MiniCPM_VQA_Simple:
                 torch_dtype=torch.bfloat16 if self.bf16_support else torch.float16,
             )
 
+
         with torch.no_grad():
             images = images.permute([0, 3, 1, 2])
             images = [ToPILImage()(img).convert("RGB") for img in images]
@@ -113,6 +119,30 @@ class MiniCPM_VQA_Simple:
                 # max_new_tokens=max_new_tokens,
                 **params,
             )
+
+            keyword_result=""
+
+            if extract_keywords:#extract_keywords
+                keyword_prompt = f"""Please extract keywords from the following text, including all occurrences of language (e.g. Chinese, English, etc.)：
+                [[[{result}]]]
+                Please list the keywords extracted, separated by commas. Make sure to include all important words, no matter what language. For English words, please keep the original case."""
+
+                keyword_msgs = [{'role': 'user', 'content': keyword_prompt}]
+                keyword_result = self.model.chat(
+                    image=None,
+                    msgs=keyword_msgs,
+                    tokenizer=self.tokenizer,
+                    sampling=True,
+                    # top_k=top_k,
+                    # top_p=top_p,
+                    temperature=temperature,
+                    # repetition_penalty=repetition_penalty,
+                    # max_new_tokens=max_new_tokens,
+                    **params,
+                )
+                print("keyword_result",keyword_result)
+
+               
             # offload model to GPU
             # self.model = self.model.to(torch.device("cpu"))
             # self.model.eval()
@@ -124,4 +154,4 @@ class MiniCPM_VQA_Simple:
                 torch.cuda.empty_cache()  # release GPU memory
                 torch.cuda.ipc_collect()
             # print(result)
-            return (result,)
+            return (result,keyword_result,)

nodes/PromptNode.py

@@ -187,7 +187,8 @@ class PromptImage:
                 }
             }
     
-    RETURN_TYPES = ()
+    RETURN_TYPES = ("STRING",)
+    RETURN_NAMES = ("json_str",)
    
     OUTPUT_NODE = True
 
@@ -202,12 +203,19 @@ class PromptImage:
         filename_prefix="mixlab_"
         filename_prefix += self.prefix_append
         full_output_folder, filename, counter, subfolder, filename_prefix = folder_paths.get_save_image_path(
-            filename_prefix, self.output_dir, images[0].shape[1], images[0].shape[0])
+            filename_prefix,self.output_dir, images[0].shape[1], images[0].shape[0])
         
+        full_output_folder=os.path.join(full_output_folder,'PromptImage')
+        subfolder='PromptImage'
+
         results = list()
 
         save_to_image=save_to_image[0]=='enable'
 
+        #保存到本地的json文件，记录图片和prompt的对应关系
+        output_images=[]
+        output_prompt=[]
+
         for index in range(len(images)):
             res=[]
             imgs=images[index]
@@ -215,24 +223,36 @@ class PromptImage:
             for image in imgs:
                 img=tensor2pil(image)
 
+                prompt_text=prompts[index]
+
                 metadata = None
                 if save_to_image:
                     metadata = PngInfo()
-                    prompt_text=prompts[index]
                     if prompt_text is not None:
                         metadata.add_text("prompt_text", prompt_text)
                     
                 file = f"{filename}_{index}_{counter:05}_.png"
-                img.save(os.path.join(full_output_folder, file), pnginfo=metadata, compress_level=self.compress_level)
+                fp=os.path.join(full_output_folder,file)
+                img.save(fp, pnginfo=metadata, compress_level=self.compress_level)
                 res.append({
                     "filename": file,
                     "subfolder": subfolder,
                     "type": self.type
                 })
+                output_images.append(fp)
+                output_prompt.append(prompt_text)
                 counter += 1
             results.append(res)
-        
-        return { "ui": { "_images": results,"prompts":prompts } }
+
+        # if save_to_image:
+        #     # 保存为本地文件
+        #     with open(os.path.join(full_output_folder,'PromptImage.json'), 'w') as file:
+        #         json.dump(output_dict, file, ensure_ascii=False, indent=4)
+
+        return { "ui": { "_images": results,"prompts":prompts },"result":(json.dumps({
+            "images":output_images,
+            "prompts":output_prompt
+        }),) }
 
 
 

nodes/SenseVoice.py

@@ -0,0 +1,226 @@
+# -*- coding:utf-8 -*-
+import logging
+import os
+import time
+
+from huggingface_hub import snapshot_download
+import torch,re
+from sensevoice.onnx.sense_voice_ort_session import SenseVoiceInferenceSession
+from sensevoice.utils.frontend import WavFrontend
+from sensevoice.utils.fsmn_vad import FSMNVad
+import comfy.utils
+import folder_paths
+
+languages = {"auto": 0, "zh": 3, "en": 4, "yue": 7, "ja": 11, "ko": 12, "nospeech": 13}
+
+# 设置环境变量
+os.environ['HF_ENDPOINT'] = 'https://hf-mirror.com'
+
+# 
+def get_model_path():
+    try:
+        return folder_paths.get_folder_paths('sense_voice')[0]
+    except:
+        return os.path.join(folder_paths.models_dir, "sense_voice")
+
+class AnyType(str):
+  """A special class that is always equal in not equal comparisons. Credit to pythongosssss"""
+
+  def __ne__(self, __value: object) -> bool:
+    return False
+
+any_type = AnyType("*")
+
+# 字幕
+def format_to_srt(channel_id, start_time_ms, end_time_ms, asr_result):
+    start_time = start_time_ms / 1000
+    end_time = end_time_ms / 1000
+    
+    def format_time(seconds):
+        hours = int(seconds // 3600)
+        minutes = int((seconds % 3600) // 60)
+        seconds = seconds % 60
+        milliseconds = int((seconds - int(seconds)) * 1000)
+        return f"{hours:02}:{minutes:02}:{int(seconds):02},{milliseconds:03}"
+    
+    start_time_str = format_time(start_time)
+    end_time_str = format_time(end_time)
+
+    pattern = r"<\|(.+?)\|><\|(.+?)\|><\|(.+?)\|><\|(.+?)\|>(.+)"
+    match = re.match(pattern,asr_result)
+    print('#format_to_srt',match,asr_result)
+    if match==None:
+        return None, None, None, None,None,start_time,end_time,None
+    lang, emotion, audio_type, itn, text = match.groups()
+     # 😊 表示高兴，😡 表示愤怒，😔 表示悲伤。对于音频事件，🎼 表示音乐，😀 表示笑声，👏 表示掌声
+    
+    srt_content = f"1\n{start_time_str} --> {end_time_str}\n{text}\n"
+    
+    logging.info(f"[Channel {channel_id}] [{start_time}s - {end_time}s] [{lang}] [{emotion}] [{audio_type}] [{itn}] {text}")
+
+    return lang, emotion, audio_type, itn,srt_content,start_time,end_time,text
+
+
+class SenseVoiceProcessor:
+    def __init__(self, download_model_path, device, num_threads, use_int8):
+
+        if not os.path.exists(download_model_path):
+            logging.info(
+                "Downloading model from huggingface hub from https://huggingface.co/lovemefan/SenseVoice-onnx"
+            )
+            logging.info(
+                "You can speed up with  `export HF_ENDPOINT=https://hf-mirror.com`"
+            )
+            snapshot_download(
+                repo_id="lovemefan/SenseVoice-onnx", local_dir=download_model_path
+            )
+
+        self.download_model_path = download_model_path
+        self.device = device
+        self.num_threads = num_threads
+        self.use_int8 = use_int8
+        self.front = WavFrontend(os.path.join(download_model_path, "am.mvn"))
+        self.model = SenseVoiceInferenceSession(
+            os.path.join(download_model_path, "embedding.npy"),
+            os.path.join(
+                download_model_path,
+                "sense-voice-encoder-int8.onnx"
+                if use_int8
+                else "sense-voice-encoder.onnx",
+            ),
+            os.path.join(download_model_path, "chn_jpn_yue_eng_ko_spectok.bpe.model"),
+            device,
+            num_threads,
+        )
+        self.vad = FSMNVad(download_model_path)
+
+    def process_audio(self, waveform, _sample_rate, language, use_itn):
+    
+        start = time.time()
+        pbar = comfy.utils.ProgressBar(waveform.shape[1])  # 进度条
+
+        results = []
+
+        for channel_id, channel_data in enumerate(waveform.T):
+            segments = self.vad.segments_offline(channel_data)
+            
+            for part in segments:
+                audio_feats = self.front.get_features(channel_data[part[0] * 16 : part[1] * 16])
+                asr_result = self.model(
+                    audio_feats[None, ...],
+                    language=languages[language],
+                    use_itn=use_itn,
+                )
+
+                lang, emotion, audio_type, itn,srt_content,start_time,end_time,text=format_to_srt(
+                    channel_id, 
+                    part[0] , 
+                    part[1], 
+                    asr_result)
+
+                if lang!=None:
+                    results.append({
+                        "language":lang, 
+                        "emotion":emotion,
+                        "audio_type":audio_type, 
+                        "itn":itn,
+                        "srt_content":srt_content,
+                        "start_time":start_time,
+                        "end_time":end_time,
+                        "text":text
+                    })
+
+            self.vad.vad.all_reset_detection()
+            pbar.update(1)  # 更新进度条
+
+        decoding_time = time.time() - start
+        logging.info(f"Decoder audio takes {decoding_time} seconds")
+        logging.info(f"The RTF is {decoding_time/(waveform.shape[1] * len(waveform) / _sample_rate)}.")
+        return results
+
+
+class SenseVoiceNode:
+
+    def __init__(self):
+        self.processor = None
+        self.download_model_path=get_model_path()
+        self.device="cpu"
+        self.num_threads = 4
+        self.use_int8 = True
+        self.language='auto'
+
+    @classmethod
+    def INPUT_TYPES(s):
+       
+        return {"required": {
+                    "audio": ("AUDIO", ), 
+                    "device": ( ['auto','cpu'], {"default": 'auto'}),
+                     "language": (list(languages.keys()), {"default": 'auto'}),# 不能直接写 languages.keys(),json.dumps会报错
+                     "num_threads":("INT",{
+                                "default":4, 
+                                "min": 1, #Minimum value
+                                "max": 32, #Maximum value
+                                "step": 1, #Slider's step
+                                "display": "number" # Cosmetic only: display as "number" or "slider"
+                                },),
+                     "use_int8":("BOOLEAN", {"default": True},),
+                     "use_itn":("BOOLEAN", {"default": True},),
+                     },
+                }
+
+    CATEGORY = "♾️Mixlab/Audio"
+
+    OUTPUT_NODE = True
+    FUNCTION = "run" 
+    
+    RETURN_TYPES = (any_type,"STRING","STRING","FLOAT",)
+    RETURN_NAMES = ("result","srt","text","total_seconds",)
+
+    def run(self,audio,device,language,num_threads,use_int8,use_itn ):
+ 
+        if device!=self.device:
+            self.device=device
+            self.processor=None
+        if language!=self.language:
+            self.language=language
+            self.processor=None
+        if num_threads!=self.num_threads:
+            self.num_threads=num_threads
+            self.processor=None
+        if use_int8!=self.use_int8:
+            self.use_int8=use_int8
+            self.processor=None
+        
+        if device=='auto' and torch.cuda.is_available():
+            self.device='cuda'
+
+        # num_threads=4
+        # use_int8=True
+
+        if self.processor==None:
+            self.processor = SenseVoiceProcessor(self.download_model_path, 
+                                                 self.device, 
+                                                 self.num_threads, 
+                                                 self.use_int8)
+
+        if 'waveform' in audio and 'sample_rate' in audio:
+            waveform = audio['waveform']
+            sample_rate = audio['sample_rate']
+            # print("Original shape:", waveform.shape)  # 打印原始形状
+            if waveform.ndim == 3 and waveform.shape[0] == 1:  # 检查是否为三维且 batch_size 为 1
+                waveform = waveform.squeeze(0)  # 移除 batch_size 维度
+            else:
+                raise ValueError("Unexpected waveform dimensions")
+
+            print("waveform.shape:", waveform.shape) 
+            total_length_seconds = waveform.shape[1] / sample_rate
+
+            waveform_numpy = waveform.numpy().transpose(1, 0)  # 转换为 (num_samples, num_channels)
+
+        results=self.processor.process_audio(waveform_numpy, sample_rate, language, use_itn)
+
+        srt_content="\n".join([s['srt_content'] for s in results])
+        text="\n".join([s['text'] for s in results])
+
+        return (results,srt_content,text,total_length_seconds,)
+ 

nodes/Utils.py

@@ -648,6 +648,49 @@ class AppInfo:
     
 
     
+class CreateJsonNode:
+    @classmethod
+    def INPUT_TYPES(s):
+        return {"required": { 
+                    "key": ("STRING",{"multiline": False,"default": "data","dynamicPrompts": False}),
+                    "value":(any_type,),
+                    "save":("BOOLEAN", {"default": True},),
+                             }, 
+                "optional":{
+                    "json_str":("STRING", {"forceInput": True,}),
+                }
+
+                }
+    
+    RETURN_TYPES = ("STRING",)
+    RETURN_NAMES = ("json_str",)
+
+    FUNCTION = "run"
+
+    CATEGORY = "♾️Mixlab/Output"
+
+    OUTPUT_NODE = True
+    INPUT_IS_LIST = False
+    # OUTPUT_IS_LIST = (True,)
+
+    def run(self,key,value,save,json_str=None):
+        data={}
+
+        data[key]=value
+
+        if json_str:
+            json_obj = json.loads(json_str)
+            data.update(json_obj)
+
+        if save:
+            # 保存为本地文件
+            with open(os.path.join(folder_paths.get_output_directory(),'data.json'), 'w') as file:
+                json.dump(data, file, ensure_ascii=False, indent=4)
+
+        return (json.dumps(data),)
+    
+
+    
 
 
 class SwitchByIndex:

nodes/Whisper.py

@@ -0,0 +1,173 @@
+import os,re
+import sys,time
+from pathlib import Path
+import torchaudio
+import hashlib
+import torch
+import folder_paths
+import comfy.utils
+
+from faster_whisper import WhisperModel
+
+class AnyType(str):
+  """A special class that is always equal in not equal comparisons. Credit to pythongosssss"""
+
+  def __ne__(self, __value: object) -> bool:
+    return False
+
+any_type = AnyType("*")
+
+def get_model_dir(m):
+    try:
+        return folder_paths.get_folder_paths(m)[0]
+    except:
+        return os.path.join(folder_paths.models_dir, m)
+
+
+
+whisper_model_path=get_model_dir('whisper')
+
+model_sizes=[
+                d for d in os.listdir(whisper_model_path) if os.path.isdir(
+                    os.path.join(whisper_model_path, d)
+                    ) and os.path.isfile(os.path.join(os.path.join(whisper_model_path, d), "config.json"))
+                    ]
+
+
+class LoadWhisperModel:
+    def __init__(self):
+        self.model = None
+        self.device="cuda" if torch.cuda.is_available() else "cpu"
+        self.model_size=model_sizes[0]
+        self.compute_type='float16'
+
+    @classmethod
+    def INPUT_TYPES(s):
+        return {"required": {
+            "model_size": (model_sizes,),
+            "device": (["auto","cpu"],),
+            "compute_type": (["float16","int8_float16","int8"],),
+                             },
+                }
+    
+    RETURN_TYPES = ("WHISPER",)
+    RETURN_NAMES = ("whisper_model",)
+
+    FUNCTION = "run"
+
+    CATEGORY = "♾️Mixlab/Audio/Whisper"
+
+    INPUT_IS_LIST = False
+    OUTPUT_IS_LIST = (False,)
+
+    def run(self,model_size,device,compute_type):
+
+        if device=="auto" and self.device!='cuda':
+            self.device="cuda" if torch.cuda.is_available() else "cpu"
+            self.model=None
+
+        if device=='cpu' and self.device!='cpu':
+            self.device="cpu"
+            self.model=None
+        
+        if model_size!= self.model_size:
+            self.model_size=model_size
+            self.model=None
+
+        if compute_type!=self.compute_type:
+            self.compute_type=compute_type
+            self.model=None
+
+        if self.model==None:
+            self.model = WhisperModel(
+                os.path.join(whisper_model_path, self.model_size), 
+                device=self.device,
+                compute_type=self.compute_type
+                )
+
+        return (self.model,)
+    
+
+class WhisperTranscribe:
+    @classmethod
+    def INPUT_TYPES(s):
+        return {"required": {
+                                "whisper_model": ("WHISPER",),
+                                "audio": ("AUDIO",),
+                             },
+                }
+    
+    RETURN_TYPES = (any_type,"STRING","STRING","FLOAT",)
+    RETURN_NAMES = ("result","srt","text","total_seconds",)
+
+    FUNCTION = "run"
+
+    CATEGORY = "♾️Mixlab/Audio/Whisper"
+
+    INPUT_IS_LIST = False
+    # OUTPUT_IS_LIST = (False,False,False,)
+    
+    def run(self,whisper_model,audio):
+
+        if 'audio_path' in audio and (not 'waveform' in audio):
+            waveform, sample_rate = torchaudio.load(audio['audio_path'])
+            waveform=waveform.mean(0)
+            total_length_seconds = waveform.shape[0] / sample_rate
+            waveform=waveform.numpy()
+            
+        elif 'waveform' in audio and 'sample_rate' in audio:
+            print("Original shape:", audio["waveform"].shape, isinstance(audio["waveform"], torch.Tensor))  # 打印原始形状
+            waveform = audio["waveform"].squeeze(0)  # Remove the added batch dimension
+            sample_rate = audio["sample_rate"]
+
+            # if audio_sf != sampling_rate:
+            #     waveform = torchaudio.functional.resample(
+            #         waveform, orig_freq=audio_sf, new_freq=sampling_rate
+            #     )
+            
+            waveform=waveform.mean(0)
+
+            total_length_seconds = waveform.shape[0] / sample_rate
+
+            waveform=waveform.numpy() #whisper_model.transcribe 旧版不支持直接传tensor，先用numpy
+        
+        segments, info = whisper_model.transcribe(waveform, beam_size=5)
+
+        print("Detected language '%s' with probability %f" % (info.language, info.language_probability))
+
+        # Function to format time for SRT
+        def format_time(seconds):
+            millis = int((seconds - int(seconds)) * 1000)
+            hours, remainder = divmod(int(seconds), 3600)
+            minutes, seconds = divmod(remainder, 60)
+            return f"{hours:02}:{minutes:02}:{seconds:02},{millis:03}"
+
+        # Prepare SRT content as a string
+        results = []
+        for i, segment in enumerate(segments):
+            start_time = format_time(segment.start)
+            end_time = format_time(segment.end)
+            srt_content = f"{i + 1}\n"
+            srt_content += f"{start_time} --> {end_time}\n"
+
+            text=segment.text.strip()
+
+            srt_content += f"{text}\n\n"
+
+            start_time=segment.start
+            end_time=segment.end
+            
+
+            results.append({ 
+                    "srt_content":srt_content,
+                    "start_time":start_time,
+                    "end_time":end_time,
+                    "text":text,
+                    "language":[info.language]
+            })
+        
+        srt_content="\n".join([s['srt_content'] for s in results])
+        text="\n".join([s['text'] for s in results])
+
+        return (results,srt_content,text,total_length_seconds,)
+

nodes/scenedetectNode.py

@@ -132,6 +132,9 @@ def split_video_by_scenes(video_path, scenes, output_path, number_of_sample_fram
     width = int(video.get(cv2.CAP_PROP_FRAME_WIDTH))
     height = int(video.get(cv2.CAP_PROP_FRAME_HEIGHT))
     
+    # 视频的总帧数
+    total_frames = int(video.get(cv2.CAP_PROP_FRAME_COUNT))
+
     # Create a list to hold the paths of the scene videos
     scenes_video = []
     keyframes = []
@@ -194,7 +197,7 @@ def split_video_by_scenes(video_path, scenes, output_path, number_of_sample_fram
     # Release the video file
     video.release()
     
-    return scenes_video, keyframes
+    return scenes_video, keyframes,total_frames
 
 
 def get_files_with_extension(directory, extension):
@@ -287,8 +290,8 @@ class ScenedetectNode_:
                      "number_of_sample_frames": ("INT", {"default": 1, "min": 1, "step": 1}), # 抽取的帧数，默认是1帧，中间帧
                      },}
 
-    RETURN_TYPES = ("SCENE_VIDEO","SCENE_", "INT",)
-    RETURN_NAMES = ("scenes_video","scenes","scene_len",)
+    RETURN_TYPES = ("SCENE_VIDEO","SCENE_", "INT","INT",)
+    RETURN_NAMES = ("scenes_video","scenes","scene_len","total_frames",)
     OUTPUT_IS_LIST = (False,False,False,)
 
     FUNCTION = "run"
@@ -310,9 +313,9 @@ class ScenedetectNode_:
         folder_path = create_folder(tp,name_without_extension)
         # print("New folder created:", folder_path)
 
-        vs_files,keyframes=split_video_by_scenes(video_path,scenes,folder_path,number_of_sample_frames)
+        vs_files,keyframes,total=split_video_by_scenes(video_path,scenes,folder_path,number_of_sample_frames)
         # print("New folder created:", vs_files)
 
-        return (vs_files,keyframes,len(scenes),)
+        return (vs_files,keyframes,len(scenes),total,)
     
 

pyproject.toml

@@ -1,7 +1,7 @@
 [project]
 name = "comfyui-mixlab-nodes"
 description = "3D, ScreenShareNode & FloatingVideoNode, SpeechRecognition & SpeechSynthesis, GPT, LoadImagesFromLocal, Layers, Other Nodes, ..."
-version = "0.41.1"
+version = "0.44.0"
 license = "MIT"
 dependencies = ["numpy", "pyOpenSSL", "watchdog", "opencv-python-headless", "matplotlib", "openai", "simple-lama-inpainting", "clip-interrogator==0.6.0", "transformers>=4.36.0", "lark-parser", "imageio-ffmpeg", "rembg[gpu]", "omegaconf==2.3.0", "Pillow>=9.5.0", "einops==0.7.0", "trimesh>=4.0.5", "huggingface-hub", "scikit-image"]
 

requirements.txt

@@ -27,4 +27,9 @@ scenedetect[opencv-headless]
 hydra-core>=1.3.2
 loralib>=0.1.2
 natsort>=8.4.0
-# simple-lama-inpainting
+
+#simple-lama-inpainting
+
+git+https://github.com/shadowcz007/SenseVoice-python.git
+
+faster_whisper

web/javascript/checkVersion_mixlab.js

@@ -3,7 +3,7 @@ import { app } from '../../../scripts/app.js'
 const repoOwner = 'shadowcz007' // 替换为仓库的所有者
 const repoName = 'comfyui-mixlab-nodes' // 替换为仓库的名称
 
-const version = 'v0.41.1'
+const version = 'v0.44.0'
 
 fetch(`https://api.github.com/repos/${repoOwner}/${repoName}/releases/latest`)
   .then(response => response.json())

web/javascript/td_background.js

@@ -6,6 +6,10 @@ window._bg_img = null
  * draws the back canvas (the one containing the background and the connections)
  * @method drawBackCanvas
  **/
+
+// 判断是否是新版的，LGraphCanvas.prototype.drawBackCanvas.toString().match('window.devicePixelRatio')
+let scale=LGraphCanvas.prototype.drawBackCanvas.toString().match('window.devicePixelRatio')?window.devicePixelRatio:1;
+
 LGraphCanvas.prototype.drawBackCanvas = function () {
   var canvas = this.bgcanvas
   if (
@@ -60,7 +64,7 @@ LGraphCanvas.prototype.drawBackCanvas = function () {
   if (!this.viewport) {
     ctx.restore()
     // ctx.setTransform(1, 0, 0, 1, 0, 0)
-    ctx.setTransform(window.devicePixelRatio, 0, 0, window.devicePixelRatio, 0, 0)
+    ctx.setTransform(scale, 0, 0, scale, 0, 0)
   }
   this.visible_links.length = 0
 

web/javascript/ui_mixlab.js

@@ -544,7 +544,7 @@ async function getCustomnodeMappings () {
     const data = (await get_nodes_map()).data
     window._nodes_maps = data
   }
-  console.log('#getCustomnodeMappings', window._nodes_maps)
+  // console.log('#getCustomnodeMappings', window._nodes_maps)
   for (let url in window._nodes_maps) {
     let n = window._nodes_maps[url]
     for (let node of n[0]) {

web/javascript/visualization.js

@@ -41,6 +41,9 @@ class Visualizer {
       overflow: 'hidden'
     })
     this.iframe.src = '/mixlab/app/' + visualSrc + '.html'
+    // this.iframe.width="300";
+    // this.iframe.height="400";
+
     console.log('#Visualizer', container, this.iframe)
     container.appendChild(this.iframe)
   }
@@ -73,7 +76,7 @@ function createVisualizer (node, inputName, typeName, inputData, app) {
     draw: function (ctx, node, widgetWidth, widgetY, widgetHeight) {
       const margin = 10
       const top_offset = 5
-      const visible = app.canvas.ds.scale > 0.5 && this.type === typeName
+      const visible = app.canvas.ds.scale > 0.3 && this.type === typeName
       const w = widgetWidth - margin * 4
       const clientRectBound = ctx.canvas.getBoundingClientRect()
       const transform = new DOMMatrix()
@@ -85,12 +88,13 @@ function createVisualizer (node, inputName, typeName, inputData, app) {
         .translateSelf(margin, margin + widgetY)
 
       Object.assign(this.visualizer.style, {
-        left: `${transform.a * margin + transform.e + 40}px`,
+        left: `${transform.a * margin + transform.e + 0}px`,
         top: `${transform.d + transform.f + top_offset}px`,
         width: `${w * transform.a}px`,
-        height: `${
-          w * transform.d - widgetHeight - margin * 15 * transform.d
-        }px`,
+        height: `${(w * transform.a * 4) / 3 - margin * 5 * transform.d}px`,
+        // height: `${
+        //   w * transform.d - widgetHeight - margin * 15 * transform.d
+        // }px`,
         position: 'absolute',
         overflow: 'hidden',
         zIndex: app.graph._nodes.indexOf(node)
@@ -137,11 +141,11 @@ function createVisualizer (node, inputName, typeName, inputData, app) {
   // Make sure visualization iframe is always inside the node when resize the node
   node.onResize = function () {
     let [w, h] = this.size
-    if (w <= 600) w = 600
-    if (h <= 500) h = 500
+    if (w <= 300) w = 300
+    if (h <= 400) h = 400
 
-    if (w > 600) {
-      h = w - 100
+    if (w > 300) {
+      h = Math.round((w * 4) / 3)
     }
 
     this.size = [w, h]
@@ -181,14 +185,14 @@ function registerVisualizer (nodeType, nodeData, nodeClassName, typeName) {
         app
       ])
 
-      this.setSize([600, 500])
+      this.setSize([300, 400])
 
       return r
     }
 
     nodeType.prototype.onExecuted = async function (message) {
       // Check if reference image and depth map are available
-      console.log("#message",message)
+      console.log('#message', message)
       if (message.reference_image && message.depth_map) {
         const params = {}
         params.reference_image = message.reference_image[0]

webApp/custom.html

@@ -2026,8 +2026,8 @@
             var iframe = document.createElement('iframe')
             iframe.src = "https://mememagic-editor.vercel.app/"
             iframe.setAttribute('frameborder', '0')
-            iframe.setAttribute('width', '500')
-            iframe.setAttribute('height', '700')
+            iframe.setAttribute('width', '700')
+            iframe.setAttribute('height', '720')
             iframe.setAttribute('allow',"clipboard-read; clipboard-write")
            
 

webApp/index.html

@@ -223,6 +223,10 @@
             margin-top: 0;
         }
 
+        .image-with-grid img {
+            margin: 0 !important;
+        }
+
         /* .card:hover {
     box-shadow: 0px 0px 10px 10px #e9fbfa;
 } */

webApp/lib/threeVisualizer.js

@@ -237,8 +237,11 @@ const sleep = (t = 1000) => {
 // 方法：旋转摄像机并拍摄图片 // 每次旋转的角度增量，转换为弧度
 async function captureImages (
   totalFrames = 20,
-  angleIncrement = THREE.MathUtils.degToRad(1.5)
+  angleIncrement = 1.5,
+  scaleFactor = 1 // 添加放大倍数参数，默认为1
 ) {
+  angleIncrement = THREE.MathUtils.degToRad(angleIncrement)
+
   // 计算场景中所有物体的中心点
   const box = new THREE.Box3().setFromObject(scene)
   const center = new THREE.Vector3()
@@ -264,6 +267,21 @@ async function captureImages (
   const startAngle = initialAngle
   //  - (angleIncrement * totalFrames) / 2
 
+  // 保存原始尺寸
+  const originalWidth = renderer.domElement.width
+  const originalHeight = renderer.domElement.height
+
+  // 调整渲染器尺寸
+  renderer.setSize(
+    originalWidth * scaleFactor,
+    originalHeight * scaleFactor,
+    false
+  )
+
+  // 调整相机的视图矩阵（如果需要）
+  camera.aspect = (originalWidth * scaleFactor) / (originalHeight * scaleFactor)
+  camera.updateProjectionMatrix()
+
   for (let i = 0; i < totalFrames; i++) {
     const angle = startAngle + i * angleIncrement
 
@@ -284,6 +302,13 @@ async function captureImages (
     await new Promise(resolve => setTimeout(resolve, 500))
   }
 
+  // 恢复渲染器尺寸
+  renderer.setSize(originalWidth, originalHeight, false)
+
+  // 恢复相机的视图矩阵
+  camera.aspect = originalWidth / originalHeight
+  camera.updateProjectionMatrix()
+
   // 恢复相机到初始位置和朝向
   camera.position.copy(initialPosition)
   camera.lookAt(initialTarget)
@@ -294,7 +319,7 @@ async function captureImages (
 async function takeScreenshot () {
   // 更新相机的矩阵，以确保其世界矩阵是最新的
   camera.updateMatrixWorld()
-  const imgs = await captureImages()
+  const imgs = await captureImages(12,3,4)
 
   // 获取当前网页的 URL
   const currentUrl = window.location.href