diff --git a/README.md b/README.md
index 49e76c44..2d8fd42e 100644
--- a/README.md
+++ b/README.md
@@ -148,33 +148,38 @@ We also hope you note that we have not verified, maintained, or updated third-pa
 To prepare the Python environment and install additional packages such as opencv, diffusers, mmcv, etc., please follow the steps below:
 
 ### Build environment
-We recommend Python 3.10 and CUDA 11.7. Set up your environment as follows:
+We recommend Python 3.10 and CUDA 11.7 or 12.x. Set up your environment as follows:
 
 ```shell
 conda create -n MuseTalk python==3.10
 conda activate MuseTalk
 ```
 
-### Install PyTorch 2.0.1
-Choose one of the following installation methods:
+### Install PyTorch
+Install PyTorch **before** other dependencies. Choose the build that matches your GPU and CUDA:
 
 ```shell
-# Option 1: Using pip
+# Option 1: CUDA 11.8 (most GPUs, e.g. RTX 20/30 series)
 pip install torch==2.0.1 torchvision==0.15.2 torchaudio==2.0.2 --index-url https://download.pytorch.org/whl/cu118
 
-# Option 2: Using conda
-conda install pytorch==2.0.1 torchvision==0.15.2 torchaudio==2.0.2 pytorch-cuda=11.8 -c pytorch -c nvidia
+# Option 2: CUDA 12.8 (RTX 50 series / Blackwell, sm_120 support)
+pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu128
+
+# Option 3: CPU only (slower inference)
+pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cpu
 ```
 
+**Note:** For NVIDIA RTX 5060 Ti / RTX 50 series (Blackwell), use the CUDA 12.8 index above. Older PyTorch builds do not include kernels for sm_120.
+
 ### Install Dependencies
-Install the remaining required packages:
+Install the remaining required packages (use `numpy<2` for opencv compatibility):
 
 ```shell
 pip install -r requirements.txt
 ```
 
 ### Install MMLab Packages
-Install the MMLab ecosystem packages:
+Install the MMLab ecosystem packages via OpenMIM:
 
 ```bash
 pip install --no-cache-dir -U openmim
@@ -184,6 +189,8 @@ mim install "mmdet==3.1.0"
 mim install "mmpose==1.1.0"
 ```
 
+**Alternative (mmcv-lite):** If no pre-built mmcv wheel is available for your PyTorch/CUDA (e.g. some Windows + CUDA 12.8 setups), you can use `mmcv-lite` instead of full mmcv. The realtime inference code includes fallbacks for NMS and ROI align when using mmcv-lite.
+
 ### Setup FFmpeg
 1. [Download](https://github.com/BtbN/FFmpeg-Builds/releases) the ffmpeg-static package
 
@@ -324,7 +331,6 @@ Important notes for real-time inference:
 2. After preparation, the avatar will generate videos using audio clips from `audio_clips`
 3. The generation process can achieve 30fps+ on an NVIDIA Tesla V100
 4. Set `preparation` to `False` for generating more videos with the same avatar
-
 For faster generation without saving images, you can use:
 ```bash
 python -m scripts.realtime_inference --inference_config configs/inference/realtime.yaml --skip_save_images
diff --git a/musetalk/utils/face_detection/detection/sfd/sfd_detector.py b/musetalk/utils/face_detection/detection/sfd/sfd_detector.py
index 8fbce152..5a92c8f3 100755
--- a/musetalk/utils/face_detection/detection/sfd/sfd_detector.py
+++ b/musetalk/utils/face_detection/detection/sfd/sfd_detector.py
@@ -21,7 +21,7 @@ def __init__(self, device, path_to_detector=os.path.join(os.path.dirname(os.path
         if not os.path.isfile(path_to_detector):
             model_weights = load_url(models_urls['s3fd'])
         else:
-            model_weights = torch.load(path_to_detector)
+            model_weights = torch.load(path_to_detector, weights_only=False)
 
         self.face_detector = s3fd()
         self.face_detector.load_state_dict(model_weights)
diff --git a/musetalk/utils/face_parsing/__init__.py b/musetalk/utils/face_parsing/__init__.py
index 09c1c02a..7801a766 100755
--- a/musetalk/utils/face_parsing/__init__.py
+++ b/musetalk/utils/face_parsing/__init__.py
@@ -62,9 +62,9 @@ def model_init(self,
         net = BiSeNet(resnet_path)
         if torch.cuda.is_available():
             net.cuda()
-            net.load_state_dict(torch.load(model_pth)) 
+            net.load_state_dict(torch.load(model_pth, weights_only=False)) 
         else:
-            net.load_state_dict(torch.load(model_pth, map_location=torch.device('cpu')))
+            net.load_state_dict(torch.load(model_pth, map_location=torch.device('cpu'), weights_only=False))
         net.eval()
         return net
 
diff --git a/musetalk/utils/face_parsing/resnet.py b/musetalk/utils/face_parsing/resnet.py
index e2e5d87e..0b831406 100755
--- a/musetalk/utils/face_parsing/resnet.py
+++ b/musetalk/utils/face_parsing/resnet.py
@@ -80,7 +80,7 @@ def forward(self, x):
         return feat8, feat16, feat32
 
     def init_weight(self, model_path):
-        state_dict = torch.load(model_path) #modelzoo.load_url(resnet18_url)
+        state_dict = torch.load(model_path, weights_only=False)  # legacy .tar format
         self_state_dict = self.state_dict()
         for k, v in state_dict.items():
             if 'fc' in k: continue
diff --git a/requirements.txt b/requirements.txt
index e87aa41d..0bea097f 100644
--- a/requirements.txt
+++ b/requirements.txt
@@ -1,6 +1,11 @@
+# Install PyTorch first (see README for GPU/CPU and RTX 50 series options):
+#   pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu128  # CUDA 12.8 (e.g. RTX 50)
+#   pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu118  # CUDA 11.8
+
 diffusers==0.30.2
 accelerate==0.28.0
-numpy==1.23.5
+# numpy<2 required for opencv-python compatibility
+numpy>=1.22,<2
 tensorflow==2.12.0
 tensorboard==2.12.0
 opencv-python==4.9.0.80
@@ -18,3 +23,4 @@ imageio[ffmpeg]
 omegaconf
 ffmpeg-python
 moviepy
+openmim
diff --git a/scripts/realtime_inference.py b/scripts/realtime_inference.py
index 579b050f..bf8e9b87 100644
--- a/scripts/realtime_inference.py
+++ b/scripts/realtime_inference.py
@@ -96,7 +96,7 @@ def init(self):
                     osmakedirs([self.avatar_path, self.full_imgs_path, self.video_out_path, self.mask_out_path])
                     self.prepare_material()
                 else:
-                    self.input_latent_list_cycle = torch.load(self.latents_out_path)
+                    self.input_latent_list_cycle = torch.load(self.latents_out_path, weights_only=False)
                     with open(self.coords_path, 'rb') as f:
                         self.coord_list_cycle = pickle.load(f)
                     input_img_list = glob.glob(os.path.join(self.full_imgs_path, '*.[jpJP][pnPN]*[gG]'))
@@ -133,7 +133,7 @@ def init(self):
                 else:
                     sys.exit()
             else:
-                self.input_latent_list_cycle = torch.load(self.latents_out_path)
+                self.input_latent_list_cycle = torch.load(self.latents_out_path, weights_only=False)
                 with open(self.coords_path, 'rb') as f:
                     self.coord_list_cycle = pickle.load(f)
                 input_img_list = glob.glob(os.path.join(self.full_imgs_path, '*.[jpJP][pnPN]*[gG]'))