fisrt commit

.gitignore

@@ -142,4 +142,6 @@ data/
 !data/models/placeholder.txt
 *.iml
 .idea/
-*.pt
+*.pt
+
+.serena

THA4_SETUP.md

@@ -0,0 +1,65 @@
+# THA4 Model Setup Guide
+
+## Required Files Structure
+
+Place your THA4 model files in `data/models/tha4/` with the following structure:
+
+```
+data/models/tha4/
+├── character_model.yaml    # Configuration file (paths to other files)
+├── character.png           # Character source image (RGBA, 512x512)
+├── face_morpher.pt         # Face morphing model weights
+└── body_morpher.pt         # Body morphing model weights
+```
+
+## Example character_model.yaml
+
+Create a file named `character_model.yaml` with the following content:
+
+```yaml
+character_image_file_name: character.png
+face_morpher_file_name: face_morpher.pt
+body_morpher_file_name: body_morpher.pt
+```
+
+**Note:** All paths in the YAML are relative to the directory containing the YAML file.
+
+## Alternative Structure
+
+You can also organize files in subdirectories:
+
+```
+data/models/tha4/
+├── my_character/
+│   ├── character_model.yaml
+│   ├── image.png
+│   ├── face_morpher.pt
+│   └── body_morpher.pt
+```
+
+Then update your YAML accordingly:
+
+```yaml
+character_image_file_name: image.png
+face_morpher_file_name: face_morpher.pt
+body_morpher_file_name: body_morpher.pt
+```
+
+## Usage in EasyVtuber
+
+1. Place your files as shown above
+2. Launch the EasyVtuber launcher
+3. Select "THA4" from the "Model Select" dropdown
+4. The system will automatically:
+   - Load `character_model.yaml`
+   - Read the PNG character image
+   - Load both PT model files
+   - Use the PNG image as the source for animation
+
+## Differences from THA3
+
+- **THA4**: Uses YAML to define character (image + models as a package)
+- **THA3**: Uses separate image file selected by user + shared model files
+
+The THA4 approach bundles character and model together, making it easier to 
+switch between different characters with their trained models.

THA4_USAGE.md

@@ -0,0 +1,127 @@
+# THA4 模型使用指南 (新版本)
+
+## 文件结构
+
+THA4 模型现在支持直接放在 `data/images/` 目录中，与 THA3 的 PNG 图片一起管理。
+
+### 所需文件
+
+对于名为 `mycharacter` 的角色，需要在 `data/images/` 中放置以下文件：
+
+```
+data/images/
+├── mycharacter.yaml                 # 配置文件
+├── mycharacter.png                  # 角色源图片 (512x512 RGBA)
+├── mycharacter_face_morpher.pt      # 面部变形模型权重
+└── mycharacter_body_morpher.pt      # 身体变形模型权重
+```
+
+### YAML 配置格式
+
+YAML 文件 (`mycharacter.yaml`) 应包含相对路径指向其他文件：
+
+```yaml
+character_image_file_name: mycharacter.png
+face_morpher_file_name: mycharacter_face_morpher.pt
+body_morpher_file_name: mycharacter_body_morpher.pt
+```
+
+**重要说明：**
+- YAML 中的所有路径都是相对于 YAML 文件所在位置的相对路径
+- 由于所有文件都在 `data/images/` 中，只需使用文件名即可
+- 角色图片必须是 512x512 像素的 RGBA 格式
+- 模型文件 (`.pt`) 是从 THA4 训练得到的 PyTorch 权重
+
+## 在 EasyVtuber 中使用
+
+### 自动检测
+
+1. 将所有 4 个文件放入 `data/images/` 目录
+2. 打开启动器 (`02A.启动器.bat`)
+3. 在 Character 下拉菜单中选择 `.yaml` 文件（例如 `mycharacter`）
+4. 程序将自动：
+   - 检测这是 THA4 模型
+   - 切换到 THA4 推理模式
+   - 从 YAML 配置加载角色图片和模型
+
+### 模型选择锁定
+
+当你选择 YAML 文件时：
+- 模型**自动锁定到 THA4**
+- 使用 YAML 中的角色图片（启动器中的 Character 选择仅用于选择 YAML）
+- 手动的模型选择将被覆盖
+
+## 示例设置
+
+假设你为名为 "alice" 的角色训练了 THA4 模型：
+
+1. 将这些文件复制到 `data/images/`：
+   ```
+   alice.yaml
+   alice.png
+   alice_face_morpher.pt
+   alice_body_morpher.pt
+   ```
+
+2. 创建 `alice.yaml`，内容为：
+   ```yaml
+   character_image_file_name: alice.png
+   face_morpher_file_name: alice_face_morpher.pt
+   body_morpher_file_name: alice_body_morpher.pt
+   ```
+
+3. 在启动器中，从 Character 下拉菜单选择 "alice"
+4. 启动 - 程序将检测到 YAML 并自动使用 THA4！
+
+## 与 THA3 的区别
+
+| 特性 | THA3 | THA4 |
+|------|------|------|
+| 角色选择 | data/images/ 中的任意 PNG | 选择 YAML 文件 |
+| 模型文件 | data/models/tha3/ 中的共享模型 | 每个角色独立的 .pt 文件 |
+| 图片加载 | 运行时选择 | 与训练模型绑定 |
+| 模型类型 | Standard/Separate, Half/Full | 单一类型 (float32) |
+| 可互换性 | 可使用任意图片 | 图片与训练模型绑定 |
+
+## 文件命名建议
+
+为了保持清晰，建议使用一致的命名规则：
+
+```
+<角色名>.yaml           # 配置文件
+<角色名>.png            # 源图片
+<角色名>_face_morpher.pt   # 面部模型
+<角色名>_body_morpher.pt   # 身体模型
+```
+
+## 故障排除
+
+**错误："THA4 character model YAML not found"**
+- 检查 `.yaml` 文件是否存在于 `data/images/` 中
+- 验证文件名是否与你选择的一致
+
+**错误："No such file or directory" 找不到 PNG 或 PT 文件**
+- 检查 YAML 中的路径是否正确
+- 确保所有 4 个文件都在同一个 `data/images/` 目录中
+- 验证文件名完全匹配（区分大小写）
+
+**黑屏 / 无输出**
+- 验证角色 PNG 是 512x512 RGBA 格式
+- 检查 .pt 模型文件是否来自正确训练的 THA4 模型
+- 确保模型文件与用于训练的角色图片匹配
+
+## 旧文件迁移
+
+如果你之前将文件放在 `data/models/tha4/` 中：
+
+1. 将所有文件移动到 `data/images/`：
+   ```
+   mv data/models/tha4/character_model.yaml data/images/mycharacter.yaml
+   mv data/models/tha4/character.png data/images/mycharacter.png
+   mv data/models/tha4/face_morpher.pt data/images/mycharacter_face_morpher.pt
+   mv data/models/tha4/body_morpher.pt data/images/mycharacter_body_morpher.pt
+   ```
+
+2. 更新 YAML 文件中的路径为新的文件名
+
+3. 在启动器中选择新的 YAML 文件名

launcher2.py

@@ -80,6 +80,9 @@ def refreshList():
     for item in sorted(os.listdir(dirPath), key=lambda x: -os.path.getmtime(os.path.join(dirPath, x))):
         if '.png' == item[-4:]:
             characterList.append(item[:-4])
+        elif '.yaml' == item[-5:] or '.yml' == item[-4:]:
+            # Support THA4 YAML files
+            characterList.append(item[:-5] if item.endswith('.yaml') else item[:-4])
 
 
 refreshList()
@@ -227,9 +230,32 @@ class LauncherPanel(wx.Panel):
                   choices=['10', '15', '20', '30', '60'])
         addOption('preset', title='Performance Preset', desc='性能预设，注意修改后会覆盖后续配置',
                   choices=['Low', 'Medium', 'High', 'Ultra', 'Custom'])
-        addOption('model_select', title='Model Select', desc='选择使用的模型\nStandard Full精度较高性能较低',
-                  choices=['Seperable Half', 'Seperable Full', 'Standard Half', 'Standard Full'],
-                  mapping=['seperable_half', 'seperable_full', 'standard_half', 'standard_full'])
+        model_select_option = addOption('model_select', title='Model Select', desc='选择使用的模型\nStandard Full精度较高性能较低',
+                  choices=['Seperable Half', 'Seperable Full', 'Standard Half', 'Standard Full', 'THA4'],
+                  mapping=['seperable_half', 'seperable_full', 'standard_half', 'standard_full', 'tha4'])
+        
+        # Add event handler for character selection to auto-lock THA4 when YAML is selected
+        def on_character_change(event):
+            character_name = self.optionDict['character'].GetValue()
+            # Check if selected character has a YAML file
+            yaml_exists = False
+            for ext in ['.yaml', '.yml']:
+                if os.path.exists(os.path.join(dirPath, character_name + ext)):
+                    yaml_exists = True
+                    break
+            
+            if yaml_exists:
+                # Auto-select THA4 model and disable the control
+                model_select_option.control.SetSelection(4)  # THA4 is index 4
+                model_select_option.control.Enable(False)
+            else:
+                # Re-enable model selection for non-YAML files
+                model_select_option.control.Enable(True)
+        
+        self.optionDict['character'].control.Bind(wx.EVT_CHOICE, on_character_change)
+        
+        # Trigger initial check
+        on_character_change(None)
         addOption('ram_cache_size', title='RAM Cache Size', desc='分配内存缓存大小\n用于存储最终运算结果',
                   choices=['Off', '1GB', '2GB', '4GB', '8GB', '16GB'],
                   mapping=['0b', '1gb', '2gb', '4gb', '8gb', '16gb'])
@@ -411,10 +437,14 @@ class LauncherPanel(wx.Panel):
                     run_args.append('4')
 
             if args['model_select'] is not None:
-                if 'seperable' in args['model_select']:
-                    run_args.append('--model_seperable')
-                if 'half' in args['model_select']:
-                    run_args.append('--model_half')
+                if 'tha4' in args['model_select']:
+                    run_args.append('--model')
+                    run_args.append('tha4')
+                else:
+                    if 'seperable' in args['model_select']:
+                        run_args.append('--model_seperable')
+                    if 'half' in args['model_select']:
+                        run_args.append('--model_half')
 
             if args['frame_rate_limit'] is not None:
                 run_args.append('--frame_rate_limit')

main.py

@@ -330,12 +330,13 @@ class MouseClientProcess(Process):
 
 
 class ModelClientProcess(Process):
-    def __init__(self, input_image):
+    def __init__(self, input_image, yaml_path=None):
         super().__init__()
         self.should_terminate = Value('b', False)
         self.updated = Value('b', False)
         self.data = None
         self.input_image = input_image
+        self.yaml_path = yaml_path
         self.output_queue = Queue()
         self.input_queue = Queue()
         self.shms = [shared_memory.SharedMemory(create=True, size=args.model_output_size * args.model_output_size * 4)
@@ -346,25 +347,36 @@ class ModelClientProcess(Process):
         self.gpu_cache_hit_ratio = Value('f', 0.0)
 
     def run(self):
-        self.model = get_core(device_id=args.device_id,
-                              use_tensorrt=args.use_tensorrt,
+        # Check if using THA4 model
+        if args.model == 'tha4':
+            from tha4_core import get_tha4_core
+            self.model = get_tha4_core(
+                device_id=args.device_id,
+                use_eyebrow=args.eyebrow,
+                interpolation_scale=args.interpolation_scale if args.use_interpolation else 1,
+                yaml_path=self.yaml_path
+            )
+        else:
+            # Use existing ezvtb_rt interface for THA3 models
+            self.model = get_core(device_id=args.device_id,
+                                  use_tensorrt=args.use_tensorrt,
 
-                              model_seperable = args.model_seperable,
-                              model_half=args.model_half, 
-                              model_cache_size=args.max_gpu_cache_len, 
-                              model_use_eyebrow=args.eyebrow,
+                                  model_seperable = args.model_seperable,
+                                  model_half=args.model_half, 
+                                  model_cache_size=args.max_gpu_cache_len, 
+                                  model_use_eyebrow=args.eyebrow,
 
-                              use_interpolation=args.use_interpolation,
-                              interpolation_scale=args.interpolation_scale,
-                              interpolation_half=args.interpolation_half,
+                                  use_interpolation=args.use_interpolation,
+                                  interpolation_scale=args.interpolation_scale,
+                                  interpolation_half=args.interpolation_half,
 
-                              cacher_quality=args.cacher_quality,
-                              cacher_ram_size=args.max_cache_len,
+                                  cacher_quality=args.cacher_quality,
+                                  cacher_ram_size=args.max_cache_len,
 
-                              use_sr=args.use_sr,
-                              sr_half=args.sr_half,
-                              sr_x4=args.sr_x4,
-                              sr_noise=args.sr_noise)
+                                  use_sr=args.use_sr,
+                                  sr_half=args.sr_half,
+                                  sr_x4=args.sr_x4,
+                                  sr_noise=args.sr_noise)
         self.model.setImage(self.input_image)
         input_pose = np.zeros((1, 45), dtype=np.float32)
 
@@ -529,26 +541,47 @@ class ModelClientProcess(Process):
 
 @torch.no_grad()
 def main():
-    img = Image.open(f"data/images/{args.character}.png")
-    img = img.convert('RGBA')
-    ow, oh = img.size
-    for i, px in enumerate(img.getdata()):
-        if px[3] <= 0:
-            y = i // ow
-            x = i % ow
-            img.putpixel((x, y), (0, 0, 0, 0))
-    if ow != 512 or oh != 512:
-        img = resize_to_512_center(img)
-    if args.alpha_clean:
-        curves = {
-            'a': [
-                (60, 0),
-                (200, 255)
-            ]
-        }
-        img = apply_color_curves(img, curves)
-    input_image = np.array(img)
-    input_image = cv2.cvtColor(input_image, cv2.COLOR_RGBA2BGRA)
+    # Check if character is a YAML file (THA4 model)
+    yaml_path = None
+    is_tha4 = False
+    
+    # Try to find YAML file
+    for ext in ['.yaml', '.yml']:
+        yaml_candidate = f"data/images/{args.character}{ext}"
+        if os.path.exists(yaml_candidate):
+            yaml_path = yaml_candidate
+            is_tha4 = True
+            print(f"Detected THA4 YAML config: {yaml_path}")
+            # Override model selection to THA4
+            args.model = 'tha4'
+            break
+    
+    # Load character image (only for non-THA4 models)
+    if not is_tha4:
+        img = Image.open(f"data/images/{args.character}.png")
+        img = img.convert('RGBA')
+        ow, oh = img.size
+        for i, px in enumerate(img.getdata()):
+            if px[3] <= 0:
+                y = i // ow
+                x = i % ow
+                img.putpixel((x, y), (0, 0, 0, 0))
+        if ow != 512 or oh != 512:
+            img = resize_to_512_center(img)
+        if args.alpha_clean:
+            curves = {
+                'a': [
+                    (60, 0),
+                    (200, 255)
+                ]
+            }
+            img = apply_color_curves(img, curves)
+        input_image = np.array(img)
+        input_image = cv2.cvtColor(input_image, cv2.COLOR_RGBA2BGRA)
+    else:
+        # For THA4, create a dummy image (will be ignored)
+        input_image = np.zeros((512, 512, 4), dtype=np.uint8)
+        print("THA4 mode: Character image will be loaded from YAML")
 
     print("Character Image Loaded:", args.character)
     cap = None
@@ -652,7 +685,7 @@ def main():
     }
 
     model_output = None
-    model_process = ModelClientProcess(input_image)
+    model_process = ModelClientProcess(input_image, yaml_path=yaml_path)
     model_process.daemon = True
     model_process.start()
     model_output_nps = [np.ndarray((args.model_output_size, args.model_output_size, 4), dtype=np.uint8,

models.py

@@ -1,5 +1,6 @@
 import time
 
+import torch
 import torch.nn as nn
 
 import tha2.poser.modes.mode_20
@@ -13,6 +14,14 @@ from args import args
 
 from collections import OrderedDict
 
+# THA4 适配器导入
+try:
+    from tha4_adapter import THA4Wrapper
+    THA4_AVAILABLE = True
+except ImportError:
+    THA4_AVAILABLE = False
+    print("Warning: THA4 adapter not available")
+
 
 class TalkingAnimeLight(nn.Module):
     def __init__(self):
@@ -171,3 +180,58 @@ class TalkingAnime(nn.Module):
         rotate_image = self.two_algo_face_rotator(x, pose_vector)[:2]
         output_image = self.combiner(rotate_image[0], rotate_image[1], pose_vector)
         return output_image
+
+
+class TalkingAnime4(nn.Module):
+    """
+    THA4 model class, compatible with THA3 TalkingAnime3 interface
+    Uses THA4's Siren architecture for inference
+    PyTorch only, float32 precision
+    """
+    def __init__(self, device):
+        super(TalkingAnime4, self).__init__()
+        if not THA4_AVAILABLE:
+            raise RuntimeError(
+                "THA4 adapter not available. Check tha4_adapter.py"
+            )
+        
+        # Create THA4 wrapper
+        self.wrapper = THA4Wrapper(device=device)
+        
+        print("THA4 model initialized (float32)")
+    
+    def forward(self, image, mouth_eye_vector, pose_vector, eyebrow_vector,
+                mouth_eye_vector_c, eyebrow_vector_c, ratio=None):
+        """
+        Forward inference, compatible with THA3 interface
+        
+        Args:
+            image: [batch, 4, 512, 512]
+            mouth_eye_vector: [batch, 27]
+            pose_vector: [batch, 6]
+            eyebrow_vector: [batch, 12]
+            mouth_eye_vector_c: compressed (for caching)
+            eyebrow_vector_c: compressed (for caching)
+            ratio: GPU cache hit ratio
+            
+        Returns:
+            output_image: [batch, 4, 512, 512]
+        """
+        if args.perf == 'model':
+            tic = time.perf_counter()
+        
+        # Call THA4 wrapper
+        output_image = self.wrapper.forward(
+            image, mouth_eye_vector, pose_vector, eyebrow_vector,
+            mouth_eye_vector_c, eyebrow_vector_c, ratio
+        )
+        
+        if args.perf == 'model':
+            print(" - tha4_inference", (time.perf_counter() - tic) * 1000)
+        
+        return output_image
+    
+    def to(self, device):
+        """Move model to specified device"""
+        self.wrapper.to(device)
+        return super().to(device)

pack_release.bat

@@ -7,57 +7,57 @@ REM ========================================================================
 
 cd /D "%~dp0"
 @REM git clean -fdx
-git submodule update --init --recursive
+@REM git submodule update --init --recursive
 
-if not exist "envs" mkdir envs
-if not exist "envs\miniconda3" mkdir envs\miniconda3
+@REM if not exist "envs" mkdir envs
+@REM if not exist "envs\miniconda3" mkdir envs\miniconda3
 
-IF not EXIST %~dp0envs\miniconda3\Scripts (
-    @RD /S /Q %~dp0envs\miniconda3
-    mkdir %~dp0envs\miniconda3
-    echo "Downloading miniconda..."
-    powershell -Command "Invoke-WebRequest -Uri 'https://repo.anaconda.com/miniconda/Miniconda3-py310_24.9.2-0-Windows-x86_64.exe' -OutFile '.\envs\miniconda3.exe' -UseBasicParsing -Headers @{'User-Agent'='Mozilla/5.0 (Windows NT 10.0; Win64; x64) AppleWebKit/537.36 (KHTML, like Gecko) Chrome/123.0.0.0 Safari/537.36'}"
+@REM IF not EXIST %~dp0envs\miniconda3\Scripts (
+@REM     @RD /S /Q %~dp0envs\miniconda3
+@REM     mkdir %~dp0envs\miniconda3
+@REM     echo "Downloading miniconda..."
+@REM     powershell -Command "Invoke-WebRequest -Uri 'https://repo.anaconda.com/miniconda/Miniconda3-py310_24.9.2-0-Windows-x86_64.exe' -OutFile '.\envs\miniconda3.exe' -UseBasicParsing -Headers @{'User-Agent'='Mozilla/5.0 (Windows NT 10.0; Win64; x64) AppleWebKit/537.36 (KHTML, like Gecko) Chrome/123.0.0.0 Safari/537.36'}"
 
-    echo "Installling minconda..."
-    start /wait "" %~dp0envs\miniconda3.exe /S /AddToPath=0 /RegisterPython=0 /InstallationType=JustMe /D=%~dp0envs\miniconda3
-    echo "Successfully install minconda"
-    del envs\miniconda3.exe
-)
+@REM     echo "Installling minconda..."
+@REM     start /wait "" %~dp0envs\miniconda3.exe /S /AddToPath=0 /RegisterPython=0 /InstallationType=JustMe /D=%~dp0envs\miniconda3
+@REM     echo "Successfully install minconda"
+@REM     del envs\miniconda3.exe
+@REM )
 
-SET PATH=%~dp0envs\miniconda3\Scripts;%PATH%
+@REM SET PATH=%~dp0envs\miniconda3\Scripts;%PATH%
 
-call activate
-call conda env list
-call conda update -y --all
-call conda create -n ezvtb_rt_venv_release -c conda-forge conda-pack python=3.10 -y
-call conda activate ezvtb_rt_venv_release
-call conda env list
+@REM call activate
+@REM call conda env list
+@REM call conda update -y --all
+@REM call conda create -n ezvtb_rt_venv_release -c conda-forge conda-pack python=3.10 -y
+@REM call conda activate ezvtb_rt_venv_release
+@REM call conda env list
 
-call conda install -y -c conda-forge cudatoolkit=12.9.1 cudnn=12.9.1
-call conda install -y nvidia/label/cuda-12.9.1::cuda-nvcc-dev_win-64
-call conda install -y pycuda -c https://mirrors.tuna.tsinghua.edu.cn/anaconda/cloud/conda-forge/ 
+@REM call conda install -y -c conda-forge cudatoolkit=12.9.1 cudnn=12.9.1
+@REM call conda install -y nvidia/label/cuda-12.9.1::cuda-nvcc-dev_win-64
+@REM call conda install -y pycuda -c https://mirrors.tuna.tsinghua.edu.cn/anaconda/cloud/conda-forge/ 
 
-call conda-pack -n ezvtb_rt_venv_release -o %~dp0envs\python_embedded --format no-archive
+@REM call conda-pack -n ezvtb_rt_venv_release -o %~dp0envs\python_embedded --format no-archive
 
 
-SET PATH=%~dp0envs\python_embedded;%~dp0envs\python_embedded\Scripts;%~dp0envs\python_embedded\Library\bin;%PATH%
+@REM SET PATH=%~dp0envs\python_embedded;%~dp0envs\python_embedded\Scripts;%~dp0envs\python_embedded\Library\bin;%PATH%
 
-call python -m pip install --upgrade pip wheel -i https://mirrors.aliyun.com/pypi/simple/
-echo yes|python -m pip install nvidia-cudnn-cu12 -i https://mirrors.aliyun.com/pypi/simple/
+@REM call python -m pip install --upgrade pip wheel -i https://mirrors.aliyun.com/pypi/simple/
+@REM echo yes|python -m pip install nvidia-cudnn-cu12 -i https://mirrors.aliyun.com/pypi/simple/
 
-echo yes|pip install tensorrt_cu12_libs==10.11.0.33 tensorrt_cu12_bindings==10.11.0.33 tensorrt==10.11.0.33 --extra-index-url https://pypi.nvidia.com
+@REM echo yes|pip install tensorrt_cu12_libs==10.11.0.33 tensorrt_cu12_bindings==10.11.0.33 tensorrt==10.11.0.33 --extra-index-url https://pypi.nvidia.com
 
-call python -m pip install -r requirements.txt --no-warn-script-location -i https://mirrors.aliyun.com/pypi/simple/
+@REM call python -m pip install -r requirements.txt --no-warn-script-location -i https://mirrors.aliyun.com/pypi/simple/
 
-if not exist "data\models" mkdir data\models
-echo "Downloading model zip file..."
-powershell -Command "(New-Object Net.WebClient).DownloadFile('https://github.com/zpeng11/ezvtuber-rt/releases/download/0.0.1/20241220.zip', '.\data\models\model.zip')"
-echo "Unzipping model zip file..."
-powershell -Command "Expand-Archive -Path '.\data\models\model.zip' -DestinationPath '.\data\models\temp' -Force; Move-Item -Path '.\data\models\temp\20241220\*' -Destination '.\data\models\' -Force; Remove-Item -Path '.\data\models\temp' -Recurse -Force"
-del .\data\models\model.zip
+@REM if not exist "data\models" mkdir data\models
+@REM echo "Downloading model zip file..."
+@REM powershell -Command "(New-Object Net.WebClient).DownloadFile('https://github.com/zpeng11/ezvtuber-rt/releases/download/0.0.1/20241220.zip', '.\data\models\model.zip')"
+@REM echo "Unzipping model zip file..."
+@REM powershell -Command "Expand-Archive -Path '.\data\models\model.zip' -DestinationPath '.\data\models\temp' -Force; Move-Item -Path '.\data\models\temp\20241220\*' -Destination '.\data\models\' -Force; Remove-Item -Path '.\data\models\temp' -Recurse -Force"
+@REM del .\data\models\model.zip
 
-@REM To verify if environment works
-call python ezvtb_rt_interface.py 
+@REM @REM To verify if environment works
+@REM call python ezvtb_rt_interface.py 
 
 @REM Clean up step
 echo "Cleaning up .trt files in data\models directory..."

tha4_adapter.py

@@ -0,0 +1,128 @@
+"""
+THA4 Adapter Module - Simplified Version
+PyTorch only, no TensorRT/ONNX support
+Adapts THA4 models to existing EasyVtuber architecture with THA3-compatible interface
+Uses CharacterModel to properly load YAML + PNG + PT files
+"""
+import sys
+import os
+import torch
+
+# Add tha4 source code path
+sys.path.insert(0, os.path.join(os.path.dirname(__file__), 'tha4', 'src'))
+
+from tha4.charmodel.character_model import CharacterModel
+
+
+def convert_tha3_pose_to_tha4(eyebrow_vector, mouth_eye_vector, pose_vector):
+    """
+    Convert THA3 pose format to THA4 format
+    
+    THA3 inputs:
+        - eyebrow_vector: [12] eyebrow parameters
+        - mouth_eye_vector: [27] facial expression parameters
+        - pose_vector: [6] head pose parameters
+        
+    THA4 input:
+        - pose: [45] unified pose vector
+          - [0:39] face pose (12 eyebrow + 27 face)
+          - [39:45] body pose (6 head rotation)
+    
+    Returns:
+        torch.Tensor [batch, 45]
+    """
+    # Concatenate: [eyebrow(12), mouth_eye(27), pose(6)] = [45]
+    face_pose = torch.cat([eyebrow_vector, mouth_eye_vector], dim=1)
+    tha4_pose = torch.cat([face_pose, pose_vector], dim=1)
+    return tha4_pose
+
+
+class THA4Wrapper:
+    """
+    THA4 Poser wrapper providing THA3 TalkingAnime3-compatible interface
+    Uses CharacterModel to load YAML config + PNG image + PT models
+    """
+    
+    def __init__(self, device: torch.device, yaml_path: str = None):
+        """
+        Initialize THA4 wrapper
+        
+        Args:
+            device: torch device (cuda/cpu)
+            yaml_path: path to character model YAML (required)
+        """
+        self.device = device
+        
+        # YAML path is required
+        if yaml_path is None:
+            raise ValueError(
+                "THA4 requires a YAML config path. "
+                "Please select a .yaml file in the Character selection."
+            )
+        
+        # Check if YAML exists
+        if not os.path.exists(yaml_path):
+            raise FileNotFoundError(
+                f"THA4 character model YAML not found: {yaml_path}\n"
+                f"Expected files in data/images/:\n"
+                f"  - <character>.yaml (config with relative paths)\n"
+                f"  - <character>.png (source image)\n"
+                f"  - <character>_face_morpher.pt (face model)\n"
+                f"  - <character>_body_morpher.pt (body model)"
+            )
+        
+        # Load CharacterModel from YAML
+        print(f"Loading THA4 CharacterModel from: {yaml_path}")
+        self.character_model = CharacterModel.load(yaml_path)
+        
+        # Get poser (lazy loads .pt files)
+        self.poser = self.character_model.get_poser(device)
+        
+        # Get character image tensor (loads .png file)
+        char_img = self.character_model.get_character_image(device)
+        
+        # Add batch dimension if needed: [4, 512, 512] -> [1, 4, 512, 512]
+        if len(char_img.shape) == 3:
+            self.character_image = char_img.unsqueeze(0)
+        else:
+            self.character_image = char_img
+        
+        print(f"THA4 model loaded successfully:")
+        print(f"  Device: {device}")
+        print(f"  Image: {self.character_model.character_image_file_name}")
+        print(f"  Face: {self.character_model.face_morpher_file_name}")
+        print(f"  Body: {self.character_model.body_morpher_file_name}")
+        
+    def forward(self, image, mouth_eye_vector, pose_vector, eyebrow_vector,
+                mouth_eye_vector_c, eyebrow_vector_c, ratio=None):
+        """
+        Forward inference, compatible with THA3 TalkingAnime3.forward()
+        
+        Args:
+            image: [batch, 4, 512, 512]
+            mouth_eye_vector: [batch, 27]
+            pose_vector: [batch, 6]
+            eyebrow_vector: [batch, 12]
+            mouth_eye_vector_c: compressed version (not used by THA4)
+            eyebrow_vector_c: compressed version (not used by THA4)
+            ratio: GPU cache hit ratio (not used by THA4)
+            
+        Returns:
+            output image [batch, 4, 512, 512]
+        """
+        # Convert pose format
+        tha4_pose = convert_tha3_pose_to_tha4(
+            eyebrow_vector, mouth_eye_vector, pose_vector
+        )
+        
+        # Call THA4 poser
+        with torch.no_grad():
+            output = self.poser.pose(image, tha4_pose, output_index=0)
+        
+        return output
+    
+    def to(self, device: torch.device):
+        """Move model to specified device"""
+        self.device = device
+        self.poser.to(device)
+        return self

tha4_core.py

@@ -0,0 +1,164 @@
+"""
+THA4 Core - Pure PyTorch implementation
+Provides interface compatible with CoreTRT/CoreORT for THA4 models
+"""
+import torch
+import numpy as np
+import cv2
+from tha4_adapter import THA4Wrapper, convert_tha3_pose_to_tha4
+
+# Import color space conversion from THA4
+import sys
+import os
+sys.path.insert(0, os.path.join(os.path.dirname(__file__), 'tha4', 'src'))
+from tha4.image_util import convert_linear_to_srgb
+
+
+class THA4Core:
+    """
+    THA4 Core class with interface compatible with CoreTRT/CoreORT
+    
+    Uses CharacterModel to load YAML + PNG + PT files properly.
+    The PNG image from YAML is used as the source image.
+    
+    Expected interface:
+    - setImage(img: np.ndarray) - NOT USED, image comes from YAML
+    - inference(pose: np.ndarray) -> List[np.ndarray] - run inference
+    """
+    
+    def __init__(self, device_id=0, use_eyebrow=True, yaml_path=None, 
+                 interpolation_scale=1):
+        """
+        Initialize THA4 Core
+        
+        Args:
+            device_id: GPU device ID
+            use_eyebrow: whether to use eyebrow parameters
+            yaml_path: path to character_model.yaml
+                      Default: 'data/models/tha4/character_model.yaml'
+            interpolation_scale: number of output frames to generate
+                      (THA4 doesn't support real interpolation, will duplicate frames)
+        """
+        device_str = f'cuda:{device_id}' if torch.cuda.is_available() else 'cpu'
+        self.device = torch.device(device_str)
+        self.use_eyebrow = use_eyebrow
+        self.interpolation_scale = interpolation_scale
+        
+        # Create THA4 wrapper (loads YAML + PNG + PT)
+        self.wrapper = THA4Wrapper(device=self.device, yaml_path=yaml_path)
+        
+        # Use character image from YAML (already loaded by wrapper)
+        self.input_image_tensor = self.wrapper.character_image
+        
+        # For compatibility (no caching in THA4 core)
+        self.cacher = None
+        
+        print(f"THA4 Core initialized on {self.device}")
+        print(f"Using character image from model (ignoring setImage calls)")
+        if interpolation_scale > 1:
+            print(f"Note: THA4 doesn't support interpolation, "
+                  f"will duplicate frames {interpolation_scale}x")
+    
+    def setImage(self, img: np.ndarray):
+        """
+        Set input character image - IGNORED for THA4
+        
+        THA4 uses the character image from the YAML/PNG files loaded
+        during initialization. This method exists only for interface
+        compatibility with CoreTRT/CoreORT.
+        
+        Args:
+            img: numpy array in BGRA format (ignored)
+        """
+        # THA4 uses character image from YAML, ignore external image
+        print("THA4: setImage() called but ignored (using image from YAML)")
+    
+    def inference(self, pose: np.ndarray) -> list:
+        """
+        Run inference
+        
+        Args:
+            pose: numpy array of shape (1, 45)
+                  [0:12] - eyebrow parameters
+                  [12:39] - face parameters  
+                  [39:45] - head pose parameters
+                  
+        Returns:
+            List of output images as numpy arrays [H, W, 4] in BGRA format, uint8
+        """
+        if self.input_image_tensor is None:
+            raise RuntimeError("Image not set. Call setImage() first.")
+        
+        # Extract pose components
+        if self.use_eyebrow:
+            eyebrow_pose = pose[:, :12]  # [1, 12]
+        else:
+            eyebrow_pose = np.zeros((1, 12), dtype=np.float32)
+            
+        face_pose = pose[:, 12:39]  # [1, 27]
+        head_pose = pose[:, 39:45]  # [1, 6]
+        
+        # Convert to torch tensors
+        eyebrow_tensor = torch.from_numpy(eyebrow_pose).to(self.device)
+        face_tensor = torch.from_numpy(face_pose).to(self.device)
+        head_tensor = torch.from_numpy(head_pose).to(self.device)
+        
+        # Call wrapper (it handles the pose conversion internally)
+        # We need to create dummy compressed versions for interface compatibility
+        eyebrow_c = eyebrow_pose[0].tolist()
+        face_c = face_pose[0].tolist()
+        
+        with torch.no_grad():
+            output_tensor = self.wrapper.forward(
+                self.input_image_tensor,
+                face_tensor,
+                head_tensor,
+                eyebrow_tensor,
+                face_c,
+                eyebrow_c,
+                ratio=None
+            )
+        
+        # THA4 outputs in linear color space, convert to sRGB
+        # Output is [batch, 4, H, W] in range [-1, 1]
+        output_image = output_tensor[0].float()  # [4, 512, 512]
+        
+        # Clip to [-1, 1] and convert to [0, 1]
+        output_image = torch.clamp((output_image + 1.0) / 2.0, 0.0, 1.0)
+        
+        # Convert linear RGB to sRGB (keep alpha as-is)
+        output_image = convert_linear_to_srgb(output_image)
+        
+        # Convert to [H, W, C] numpy
+        output_np = output_image.permute(1, 2, 0).cpu().numpy()
+        
+        # Convert to uint8
+        output_np = (output_np * 255.0).astype(np.uint8)
+        
+        # Convert RGBA to BGRA for output
+        output_bgra = cv2.cvtColor(output_np, cv2.COLOR_RGBA2BGRA)
+        
+        # Return as list with proper number of frames
+        # THA4 doesn't support real interpolation, so duplicate frames if needed
+        return [output_bgra] * self.interpolation_scale
+
+
+def get_tha4_core(device_id=0, use_eyebrow=True, yaml_path=None,
+                  interpolation_scale=1):
+    """
+    Factory function to create THA4 core
+    
+    Args:
+        device_id: GPU device ID
+        use_eyebrow: whether to use eyebrow control
+        yaml_path: path to character_model.yaml
+                  Default: 'data/models/tha4/character_model.yaml'
+        interpolation_scale: number of frames (for compatibility with interpolation)
+        
+    Returns:
+        THA4Core instance
+    """
+    return THA4Core(device_id=device_id,
+                    use_eyebrow=use_eyebrow,
+                    yaml_path=yaml_path,
+                    interpolation_scale=interpolation_scale)