diff --git a/README.md b/README.md
index 7a68a2d..a948af2 100644
--- a/README.md
+++ b/README.md
@@ -49,13 +49,13 @@
 
 
 ## 📮 Update
-- [2026.03] Release inference codes and gradio demo.
+- [2026.03] Release inference codes, evaluation codes, and gradio demo.
 - [2025.12] This repo is created.
 
 
 ## 🏄🏻‍♀️ TODO
 - [x] Release inference codes and gradio demo. 
-- [ ] Release evaluation codes.
+- [x] Release evaluation codes.
 - [ ] Release training codes for video matting model.
 - [ ] Release checkpoint and training codes for quality evaluator model.
 - [ ] Release real-world video matting dataset **VMReal**.
@@ -138,6 +138,8 @@ By launching, an interactive interface will appear as follow.
 
 ![overall_teaser](assets/teaser_demo.gif)
 
+## 📊 Evaluation
+Please refer to the [evaluation documentation](docs/EVAL.md) for details.
 
 ## 🛠️ Data Pipeline
 ![data_pipeline](assets/data_pipeline.jpg)
diff --git a/docs/EVAL.md b/docs/EVAL.md
new file mode 100644
index 0000000..224df78
--- /dev/null
+++ b/docs/EVAL.md
@@ -0,0 +1,153 @@
+# Evaluation Documentation
+
+This document shows the way to reproduce the quantitative results in our paper on two synthetic benchmarks ([YouTubeMatte](https://github.com/pq-yang/MatAnyone/tree/main?tab=readme-ov-file#youtubematte-dataset) and [VideoMatte](https://github.com/PeterL1n/RobustVideoMatting/blob/master/documentation/training.md#evaluation)) and one real benchmark ([CRGNN](https://github.com/TiantianWang/VideoMatting-CRGNN)).
+
+## YouTubeMatte
+
+**📦 We provide the inference results with MatAnyone 2 on the YouTubeMatte benchmark [here](https://drive.google.com/drive/folders/1fgPAx4pRGyxGIYW4NeBevDM8PpA0TG9I?usp=sharing).**
+
+### Preparation
+* [YouTubeMatte.zip (6.24G)](https://drive.google.com/file/d/1IEH0RaimT_hSp38AWF6wuwNJzzNSHpJ4/view?usp=drive_link)
+* [YouTubeMatte_first_frame_seg_mask.zip (310K)](https://drive.google.com/file/d/1Zpa7SB7VZmkvRDiehVC-c_0dmFWXdfzK/view?usp=drive_linkk)
+
+To run the inference scripts, your files should be arranged as:
+```
+data
+   |- YouTubeMatte_first_frame_seg_mask   # for inference only
+   |- YouTubeMatte
+        |- youtubematte_512x288
+        |- youtubematte_1920x1080
+```
+
+### Batch Inference
+Empirically, for low-resolution (`youtubematte_512x288`) and high-resolution (`youtubematte_1920x1080`) data, we set **different** hyperparameter values for `--warmup`, `--erode_kernel`, and `--dilate_kernel`.
+
+```shell
+# lr: youtubematte_512x288
+bash evaluation/infer_batch_lr_yt.sh
+
+# hr: youtubematte_1920x1080
+bash evaluation/infer_batch_hr_yt.sh
+```
+
+### Evaluation
+To run the evaluation scripts, your files should be arranged as:
+
+```
+data
+   |- YouTubeMatte
+        |- youtubematte_512x288
+        |- youtubematte_1920x1080
+
+   |- results
+        |- youtubematte_512x288
+        |- youtubematte_1920x1080
+```
+
+```shell
+# lr: youtubematte_512x288
+python evaluation/eval_lr.py \
+    --pred-dir ./data/results/youtubematte_512x288 \
+    --true-dir ./data/YouTubeMatte/youtubematte_512x288 
+
+# hr: youtubematte_1920x1080
+python evaluation/eval_hr.py \
+    --pred-dir ./data/results/youtubematte_1920x1080 \
+    --true-dir ./data/YouTubeMatte/youtubematte_1920x1080 
+```
+
+## VideoMatte
+
+**📦 We provide the inference results with MatAnyone 2 on the VideoMatte benchmark [here](https://drive.google.com/drive/folders/12QM-_kyerE1tQfINoR17zYIFtrcwCK04?usp=sharing).**
+
+### Preparation
+* [videomatte_512x512.tar (1.8G)](https://robustvideomatting.blob.core.windows.net/eval/videomatte_512x288.tar)
+* [videomatte_1920x1080.tar (2.2G)](https://robustvideomatting.blob.core.windows.net/eval/videomatte_1920x1080.tar)
+* [VideoMatte_first_frame_seg_mask.zip (416K)](https://drive.google.com/file/d/1kN5gX4NAEa4HG-k2ir8kPcEp_18DbDHt/view?usp=drive_link)
+
+To run the inference scripts, your files should be arranged as:
+```
+data
+   |- VideoMatte_first_frame_seg_mask   # for inference only
+   |- VideoMatte
+        |- videomatte_512x288
+        |- videomatte_1920x1080
+```
+
+### Batch Inference
+Empirically, for low-resolution (`videomatte_512x288`) and high-resolution (`videomatte_1920x1080`) data, we set **different** hyperparameter values for `--warmup`, `--erode_kernel`, and `--dilate_kernel`.
+
+```shell
+# lr: videomatte_512x288
+bash evaluation/infer_batch_lr_vm.sh
+
+# hr: videomatte_1920x1080
+bash evaluation/infer_batch_hr_vm.sh
+```
+
+### Evaluation
+To run the evaluation scripts, your files should be arranged as:
+
+```
+data
+   |- VideoMatte
+        |- videomatte_512x288
+        |- videomatte_1920x1080
+
+   |- results
+        |- videomatte_512x288
+        |- videomatte_512x288
+```
+
+```shell
+# lr: videomatte_512x288
+python evaluation/eval_lr.py \
+    --pred-dir ./data/results/videomatte_512x288 \
+    --true-dir ./data/VideoMatte/videomatte_512x288 
+
+# hr: videomatte_1920x1080
+python evaluation/eval_hr.py \
+    --pred-dir ./data/results/videomatte_1920x1080 \
+    --true-dir ./data/VideoMatte/videomatte_1920x1080 
+```
+
+## CRGNN
+
+**📦 We provide the inference results with MatAnyone 2 on the CRGNN benchmark [here](https://drive.google.com/file/d/1JJyE4uPymEcijNa1Ok8ME_BY6oxJ3EXJ/view?usp=sharing).**
+
+### Preparation
+* [real_human_data](https://www.dropbox.com/sh/23uvsue5we7e7b5/AAB4GSSWIaKiSouvN3wuWiwWa?dl=0)
+* [CRGNN_first_frame_seg_mask.zip (151K)](https://drive.google.com/file/d/1cDSf1kO_tdWy-q3CuX4IfuZ16ZYEER-d/view?usp=sharing)
+
+To run the inference scripts, your files should be arranged as:
+```
+data
+   |- crgnn   
+     |- alpha
+     |- image_allframe
+     |- mask              # first frame seg mask
+```
+
+### Batch Inference
+
+```shell
+bash evaluation/infer_batch_crgnn.sh
+```
+
+### Evaluation
+To run the evaluation scripts, your files should be arranged as:
+
+```
+data
+   |- crgnn
+        |- alpha
+
+   |- results
+        |- crgnn
+```
+
+```shell
+python evaluation/eval_crgnn.py \
+    --pred-dir ./data/results/crgnn \
+    --true-dir ./data/crgnn/alpha 
+```
diff --git a/evaluation/eval_crgnn.py b/evaluation/eval_crgnn.py
new file mode 100644
index 0000000..3797b0f
--- /dev/null
+++ b/evaluation/eval_crgnn.py
@@ -0,0 +1,242 @@
+# This file is modified based on `evaluate_hr.py` from [RobustVideoMatting](https://github.com/PeterL1n/RobustVideoMatting)
+# Changes:
+# - Adapted for CRGNN dataset
+# - Supported metrics: pha_mad, pha_mse, pha_grad, pha_dtssd
+
+"""
+HR (High-Resolution) evaluation. We found using numpy is very slow for high resolution, so we moved it to PyTorch using CUDA.
+
+Note, the script only does evaluation. You will need to first inference yourself and save the results to disk
+Expected directory format for both prediction and ground-truth is:
+
+    crgnn
+        ├── video1
+          ├── pha
+            ├── 0000.png
+        ├── video2
+          ├── pha
+            ├── 0000.png
+
+Prediction must have the exact file structure and file name as the ground-truth,
+except for the file extension.
+
+Example usage:
+
+python evaluation/eval_crgnn.py \
+    --pred-dir ./data/results/crgnn \
+    --true-dir ./data/crgnn/alpha 
+
+An excel sheet with evaluation results will be written to "PATH_TO_PREDICTIONS/crgnn.xlsx"
+"""
+
+
+import argparse
+import os
+import cv2
+import kornia
+import numpy as np
+import xlsxwriter
+import torch
+from concurrent.futures import ThreadPoolExecutor
+from tqdm import tqdm
+
+
+class Evaluator:
+    def __init__(self):
+        self.parse_args()
+        self.init_metrics()
+        self.evaluate()
+        self.write_excel()
+        
+    def parse_args(self):
+        parser = argparse.ArgumentParser()
+        parser.add_argument('--pred-dir', type=str, required=True)
+        parser.add_argument('--true-dir', type=str, required=True)
+        parser.add_argument('--num-workers', type=int, default=48)
+        parser.add_argument('--metrics', type=str, nargs='+', default=[
+            'pha_mad', 'pha_mse', 'pha_grad', 'pha_dtssd'])
+        self.args = parser.parse_args()
+        
+    def init_metrics(self):
+        self.mad = MetricMAD()
+        self.mse = MetricMSE()
+        self.grad = MetricGRAD()
+        self.dtssd = MetricDTSSD()
+        
+    def evaluate(self):
+        tasks = []
+        position = 0
+        
+        with ThreadPoolExecutor(max_workers=self.args.num_workers) as executor:
+            for video_folder in sorted(os.listdir(self.args.pred_dir)):
+                pred_video_path = os.path.join(self.args.pred_dir, video_folder)
+                true_video_path = os.path.join(self.args.true_dir, video_folder)
+                if os.path.isdir(pred_video_path) and os.path.isdir(true_video_path):
+                    future = executor.submit(self.evaluate_worker, video_folder, position)
+                    tasks.append((video_folder, future))
+                    position += 1
+                    
+        self.results = [(video_folder, future.result()) for video_folder, future in tasks]
+        
+    def write_excel(self):
+        workbook = xlsxwriter.Workbook(os.path.join(self.args.pred_dir, f'{os.path.basename(self.args.pred_dir)}.xlsx'))
+        summarysheet = workbook.add_worksheet('summary')
+        metricsheets = [workbook.add_worksheet(metric) for metric in self.results[0][-1].keys()]
+        
+        for i, metric in enumerate(self.results[0][-1].keys()):
+            summarysheet.write(i, 0, metric)
+            summarysheet.write(i, 1, f'={metric}!B2')
+        
+        for row, (video_folder, metrics) in enumerate(self.results):
+            for metricsheet, metric in zip(metricsheets, metrics.values()):
+                # Write the header
+                if row == 0:
+                    metricsheet.write(1, 0, 'Average')
+                    metricsheet.write(1, 1, f'=AVERAGE(C2:ZZ2)')
+                    for col in range(len(metric)):
+                        metricsheet.write(0, col + 2, col)
+                        colname = xlsxwriter.utility.xl_col_to_name(col + 2)
+                        metricsheet.write(1, col + 2, f'=AVERAGE({colname}3:{colname}9999)')
+                        
+                metricsheet.write(row + 2, 0, video_folder)
+                metricsheet.write_row(row + 2, 1, metric)
+        
+        workbook.close()
+
+    def evaluate_worker(self, video_folder, position):
+        pred_pha_dir = os.path.join(self.args.pred_dir, video_folder, 'pha')
+        true_pha_dir = os.path.join(self.args.true_dir, video_folder)
+        
+        # Get sorted filenames from both directories
+        pred_framenames = sorted([f for f in os.listdir(pred_pha_dir) if os.path.isfile(os.path.join(pred_pha_dir, f))])
+        true_framenames = sorted([f for f in os.listdir(true_pha_dir) if os.path.isfile(os.path.join(true_pha_dir, f))])
+        
+        # true-dir contains frames at every 10 frames (frame 1, 11, 21, ...)
+        # pred-dir contains all frames (frame 0, 1, 2, 3, ...)
+        # Match: true frame i (which is video frame i*10) corresponds to pred frame i*10
+        num_true_frames = len(true_framenames)
+        num_pred_frames = len(pred_framenames)
+        
+        if num_true_frames == 0:
+            print(f'Warning: {video_folder} has no true frames')
+            return {metric_name : [] for metric_name in self.args.metrics}
+        
+        print(f'Info: {video_folder} has {num_pred_frames} pred frames and {num_true_frames} true frames (every 10 frames)')
+        print(f'  First few pred files: {pred_framenames[:5]}')
+        print(f'  First few true files: {true_framenames[:5]}')
+        
+        metrics = {metric_name : [] for metric_name in self.args.metrics}
+        
+        pred_pha_tm1 = None
+        true_pha_tm1 = None
+        
+        for true_idx in tqdm(range(num_true_frames), desc=video_folder, position=position, dynamic_ncols=True):
+            # true frame at index true_idx corresponds to video frame true_idx * 10
+            # pred frame at index pred_idx = true_idx * 10
+            pred_idx = true_idx * 10
+            
+            if pred_idx >= num_pred_frames:
+                print(f'Warning: {video_folder} true frame {true_idx} (video frame {pred_idx}) exceeds pred frames ({num_pred_frames})')
+                break
+            
+            pred_pha_path = os.path.join(pred_pha_dir, pred_framenames[pred_idx])
+            true_pha_path = os.path.join(true_pha_dir, true_framenames[true_idx])
+            
+            # Print the matched pair
+            print(f'  [{video_folder}] Pair {true_idx}: pred[{pred_idx}]="{pred_framenames[pred_idx]}" <-> true[{true_idx}]="{true_framenames[true_idx]}"')
+            
+            pred_pha = cv2.imread(pred_pha_path, cv2.IMREAD_GRAYSCALE)
+            true_pha = cv2.imread(true_pha_path, cv2.IMREAD_GRAYSCALE)
+            
+            if pred_pha is None or true_pha is None:
+                print(f'Warning: Failed to read image at true_idx {true_idx} (pred_idx {pred_idx}) in {video_folder}')
+                continue
+            
+            true_pha = torch.from_numpy(true_pha).cuda(non_blocking=True).float().div_(255)
+            pred_pha = torch.from_numpy(pred_pha).cuda(non_blocking=True).float().div_(255)
+            
+            if 'pha_mad' in self.args.metrics:
+                metrics['pha_mad'].append(self.mad(pred_pha, true_pha))
+            if 'pha_mse' in self.args.metrics:
+                metrics['pha_mse'].append(self.mse(pred_pha, true_pha))
+            if 'pha_grad' in self.args.metrics:
+                metrics['pha_grad'].append(self.grad(pred_pha, true_pha))
+            if 'pha_conn' in self.args.metrics:
+                metrics['pha_conn'].append(self.conn(pred_pha, true_pha))
+            if 'pha_dtssd' in self.args.metrics:
+                if true_idx == 0:
+                    metrics['pha_dtssd'].append(0)
+                else:
+                    metrics['pha_dtssd'].append(self.dtssd(pred_pha, pred_pha_tm1, true_pha, true_pha_tm1))
+                    
+            pred_pha_tm1 = pred_pha
+            true_pha_tm1 = true_pha
+            
+
+        return metrics
+
+    
+class MetricMAD:
+    def __call__(self, pred, true):
+        return (pred - true).abs_().mean() * 1e3
+
+
+class MetricMSE:
+    def __call__(self, pred, true):
+        return ((pred - true) ** 2).mean() * 1e3
+
+
+class MetricGRAD:
+    def __init__(self, sigma=1.4):
+        self.filter_x, self.filter_y = self.gauss_filter(sigma)
+        self.filter_x = torch.from_numpy(self.filter_x).unsqueeze(0).cuda()
+        self.filter_y = torch.from_numpy(self.filter_y).unsqueeze(0).cuda()
+    
+    def __call__(self, pred, true):
+        true_grad = self.gauss_gradient(true)
+        pred_grad = self.gauss_gradient(pred)
+        return ((true_grad - pred_grad) ** 2).sum() / 1000
+    
+    def gauss_gradient(self, img):
+        img_filtered_x = kornia.filters.filter2d(img[None, None, :, :], self.filter_x, border_type='replicate')[0, 0]
+        img_filtered_y = kornia.filters.filter2d(img[None, None, :, :], self.filter_y, border_type='replicate')[0, 0]
+        return (img_filtered_x**2 + img_filtered_y**2).sqrt()
+    
+    @staticmethod
+    def gauss_filter(sigma, epsilon=1e-2):
+        half_size = np.ceil(sigma * np.sqrt(-2 * np.log(np.sqrt(2 * np.pi) * sigma * epsilon)))
+        size = int(2 * half_size + 1)
+
+        # create filter in x axis
+        filter_x = np.zeros((size, size))
+        for i in range(size):
+            for j in range(size):
+                filter_x[i, j] = MetricGRAD.gaussian(i - half_size, sigma) * MetricGRAD.dgaussian(
+                    j - half_size, sigma)
+
+        # normalize filter
+        norm = np.sqrt((filter_x**2).sum())
+        filter_x = filter_x / norm
+        filter_y = np.transpose(filter_x)
+
+        return filter_x, filter_y
+        
+    @staticmethod
+    def gaussian(x, sigma):
+        return np.exp(-x**2 / (2 * sigma**2)) / (sigma * np.sqrt(2 * np.pi))
+    
+    @staticmethod
+    def dgaussian(x, sigma):
+        return -x * MetricGRAD.gaussian(x, sigma) / sigma**2
+
+
+class MetricDTSSD:
+    def __call__(self, pred_t, pred_tm1, true_t, true_tm1):
+        dtSSD = ((pred_t - pred_tm1) - (true_t - true_tm1)) ** 2
+        dtSSD = dtSSD.sum() / true_t.numel()
+        dtSSD = dtSSD.sqrt()
+        return dtSSD * 1e2
+
+
+if __name__ == '__main__':
+    Evaluator()
\ No newline at end of file
diff --git a/evaluation/eval_hr.py b/evaluation/eval_hr.py
new file mode 100644
index 0000000..37a3e46
--- /dev/null
+++ b/evaluation/eval_hr.py
@@ -0,0 +1,213 @@
+# This file is modified based on `evaluate_hr.py` from [RobustVideoMatting](https://github.com/PeterL1n/RobustVideoMatting)
+# Changes:
+# - Adapted for YouTubeMatte/VideoMatte dataset
+# - Supported metrics: pha_mad, pha_mse, pha_grad, pha_dtssd
+
+"""
+HR (High-Resolution) evaluation. We found using numpy is very slow for high resolution, so we moved it to PyTorch using CUDA.
+
+Note, the script only does evaluation. You will need to first inference yourself and save the results to disk
+Expected directory format for both prediction and ground-truth is (same for VideoMatte):
+
+    youtubematte_1920x1080
+        ├── youtubematte_motion
+          ├── 0000
+            ├── pha
+              ├── 0000.png
+        ├── youtubematte_static
+          ├── 0000
+            ├── pha
+              ├── 0000.png
+
+Prediction must have the exact file structure and file name as the ground-truth,
+except for the file extension.
+
+Example usage:
+
+[YouTubeMatte]
+python evaluation/eval_hr.py \
+    --pred-dir ./data/results/youtubematte_1920x1080 \
+    --true-dir ./data/YouTubeMatte/youtubematte_1920x1080 
+
+[VideoMatte]
+python evaluation/eval_hr.py \
+    --pred-dir ./data/results/videomatte_1920x1080 \
+    --true-dir ./data/VideoMatte/videomatte_1920x1080 
+
+An excel sheet with evaluation results will be written to "PATH_TO_PREDICTIONS/xxxmatte_1920x1080.xlsx"
+"""
+
+
+import argparse
+import os
+import cv2
+import kornia
+import numpy as np
+import xlsxwriter
+import torch
+from concurrent.futures import ThreadPoolExecutor
+from tqdm import tqdm
+
+
+class Evaluator:
+    def __init__(self):
+        self.parse_args()
+        self.init_metrics()
+        self.evaluate()
+        self.write_excel()
+        
+    def parse_args(self):
+        parser = argparse.ArgumentParser()
+        parser.add_argument('--pred-dir', type=str, required=True)
+        parser.add_argument('--true-dir', type=str, required=True)
+        parser.add_argument('--num-workers', type=int, default=48)
+        parser.add_argument('--metrics', type=str, nargs='+', default=[
+            'pha_mad', 'pha_mse', 'pha_grad', 'pha_dtssd'])
+        self.args = parser.parse_args()
+        
+    def init_metrics(self):
+        self.mad = MetricMAD()
+        self.mse = MetricMSE()
+        self.grad = MetricGRAD()
+        self.dtssd = MetricDTSSD()
+        
+    def evaluate(self):
+        tasks = []
+        position = 0
+        
+        with ThreadPoolExecutor(max_workers=self.args.num_workers) as executor:
+            for dataset in sorted(os.listdir(self.args.pred_dir)):
+                if os.path.isdir(os.path.join(self.args.pred_dir, dataset)):
+                    for clip in sorted(os.listdir(os.path.join(self.args.pred_dir, dataset))):
+                        future = executor.submit(self.evaluate_worker, dataset, clip, position)
+                        tasks.append((dataset, clip, future))
+                        position += 1
+                    
+        self.results = [(dataset, clip, future.result()) for dataset, clip, future in tasks]
+        
+    def write_excel(self):
+        workbook = xlsxwriter.Workbook(os.path.join(self.args.pred_dir, f'{os.path.basename(self.args.pred_dir)}.xlsx'))
+        summarysheet = workbook.add_worksheet('summary')
+        metricsheets = [workbook.add_worksheet(metric) for metric in self.results[0][-1].keys()]
+        
+        for i, metric in enumerate(self.results[0][-1].keys()):
+            summarysheet.write(i, 0, metric)
+            summarysheet.write(i, 1, f'={metric}!B2')
+        
+        for row, (dataset, clip, metrics) in enumerate(self.results):
+            for metricsheet, metric in zip(metricsheets, metrics.values()):
+                # Write the header
+                if row == 0:
+                    metricsheet.write(1, 0, 'Average')
+                    metricsheet.write(1, 1, f'=AVERAGE(C2:ZZ2)')
+                    for col in range(len(metric)):
+                        metricsheet.write(0, col + 2, col)
+                        colname = xlsxwriter.utility.xl_col_to_name(col + 2)
+                        metricsheet.write(1, col + 2, f'=AVERAGE({colname}3:{colname}9999)')
+                        
+                metricsheet.write(row + 2, 0, dataset)
+                metricsheet.write(row + 2, 1, clip)
+                metricsheet.write_row(row + 2, 2, metric)
+        
+        workbook.close()
+
+    def evaluate_worker(self, dataset, clip, position):
+        framenames = sorted(os.listdir(os.path.join(self.args.pred_dir, dataset, clip, 'pha')))
+        
+        metrics = {metric_name : [] for metric_name in self.args.metrics}
+        
+        pred_pha_tm1 = None
+        true_pha_tm1 = None
+        
+        for i, framename in enumerate(tqdm(framenames, desc=f'{dataset} {clip}', position=position, dynamic_ncols=True)):
+            true_pha = cv2.imread(os.path.join(self.args.true_dir, dataset, clip, 'pha', framename[:-4]+".jpg"), cv2.IMREAD_GRAYSCALE)
+            pred_pha = cv2.imread(os.path.join(self.args.pred_dir, dataset, clip, 'pha', framename), cv2.IMREAD_GRAYSCALE)
+            
+            true_pha = torch.from_numpy(true_pha).cuda(non_blocking=True).float().div_(255)
+            pred_pha = torch.from_numpy(pred_pha).cuda(non_blocking=True).float().div_(255)
+            
+            if 'pha_mad' in self.args.metrics:
+                metrics['pha_mad'].append(self.mad(pred_pha, true_pha))
+            if 'pha_mse' in self.args.metrics:
+                metrics['pha_mse'].append(self.mse(pred_pha, true_pha))
+            if 'pha_grad' in self.args.metrics:
+                metrics['pha_grad'].append(self.grad(pred_pha, true_pha))
+            if 'pha_conn' in self.args.metrics:
+                metrics['pha_conn'].append(self.conn(pred_pha, true_pha))
+            if 'pha_dtssd' in self.args.metrics:
+                if i == 0:
+                    metrics['pha_dtssd'].append(0)
+                else:
+                    metrics['pha_dtssd'].append(self.dtssd(pred_pha, pred_pha_tm1, true_pha, true_pha_tm1))
+                    
+            pred_pha_tm1 = pred_pha
+            true_pha_tm1 = true_pha
+            
+
+        return metrics
+
+    
+class MetricMAD:
+    def __call__(self, pred, true):
+        return (pred - true).abs_().mean() * 1e3
+
+
+class MetricMSE:
+    def __call__(self, pred, true):
+        return ((pred - true) ** 2).mean() * 1e3
+
+
+class MetricGRAD:
+    def __init__(self, sigma=1.4):
+        self.filter_x, self.filter_y = self.gauss_filter(sigma)
+        self.filter_x = torch.from_numpy(self.filter_x).unsqueeze(0).cuda()
+        self.filter_y = torch.from_numpy(self.filter_y).unsqueeze(0).cuda()
+    
+    def __call__(self, pred, true):
+        true_grad = self.gauss_gradient(true)
+        pred_grad = self.gauss_gradient(pred)
+        return ((true_grad - pred_grad) ** 2).sum() / 1000
+    
+    def gauss_gradient(self, img):
+        img_filtered_x = kornia.filters.filter2d(img[None, None, :, :], self.filter_x, border_type='replicate')[0, 0]
+        img_filtered_y = kornia.filters.filter2d(img[None, None, :, :], self.filter_y, border_type='replicate')[0, 0]
+        return (img_filtered_x**2 + img_filtered_y**2).sqrt()
+    
+    @staticmethod
+    def gauss_filter(sigma, epsilon=1e-2):
+        half_size = np.ceil(sigma * np.sqrt(-2 * np.log(np.sqrt(2 * np.pi) * sigma * epsilon)))
+        size = int(2 * half_size + 1)
+
+        # create filter in x axis
+        filter_x = np.zeros((size, size))
+        for i in range(size):
+            for j in range(size):
+                filter_x[i, j] = MetricGRAD.gaussian(i - half_size, sigma) * MetricGRAD.dgaussian(
+                    j - half_size, sigma)
+
+        # normalize filter
+        norm = np.sqrt((filter_x**2).sum())
+        filter_x = filter_x / norm
+        filter_y = np.transpose(filter_x)
+
+        return filter_x, filter_y
+        
+    @staticmethod
+    def gaussian(x, sigma):
+        return np.exp(-x**2 / (2 * sigma**2)) / (sigma * np.sqrt(2 * np.pi))
+    
+    @staticmethod
+    def dgaussian(x, sigma):
+        return -x * MetricGRAD.gaussian(x, sigma) / sigma**2
+
+
+class MetricDTSSD:
+    def __call__(self, pred_t, pred_tm1, true_t, true_tm1):
+        dtSSD = ((pred_t - pred_tm1) - (true_t - true_tm1)) ** 2
+        dtSSD = dtSSD.sum() / true_t.numel()
+        dtSSD = dtSSD.sqrt()
+        return dtSSD * 1e2
+
+
+if __name__ == '__main__':
+    Evaluator()
\ No newline at end of file
diff --git a/evaluation/eval_lr.py b/evaluation/eval_lr.py
new file mode 100644
index 0000000..6531c0a
--- /dev/null
+++ b/evaluation/eval_lr.py
@@ -0,0 +1,252 @@
+# This file is modified based on `evaluate_lr.py` from [RobustVideoMatting](https://github.com/PeterL1n/RobustVideoMatting)
+# Changes:
+# - Adapted for YouTubeMatte/VideoMatte dataset
+# - Supported metrics: pha_mad, pha_mse, pha_grad, pha_dtssd, pha_conn
+
+"""
+LR (Low-Resolution) evaluation.
+
+Note, the script only does evaluation. You will need to first inference yourself and save the results to disk
+Expected directory format for both prediction and ground-truth is (same for VideoMatte):
+
+    youtubematte_512x288
+        ├── youtubematte_motion
+          ├── 0000
+            ├── pha
+              ├── 0000.png
+        ├── youtubematte_static
+          ├── 0000
+            ├── pha
+              ├── 0000.png
+
+Prediction must have the exact file structure and file name as the ground-truth,
+meaning that if the ground-truth is png/jpg, prediction should be png/jpg.
+
+Example usage:
+
+[YouTubeMatte]
+python evaluation/eval_lr.py \
+    --pred-dir ./data/results/youtubematte_512x288 \
+    --true-dir ./data/YouTubeMatte/youtubematte_512x288 
+
+[VideoMatte]
+python evaluation/eval_lr.py \
+    --pred-dir ./data/results/videomatte_512x288 \
+    --true-dir ./data/VideoMatte/videomatte_512x288 
+
+An excel sheet with evaluation results will be written to "PATH_TO_PREDICTIONS/xxxmatte_512x288.xlsx"
+"""
+
+
+import argparse
+import os
+import cv2
+import numpy as np
+import xlsxwriter
+from concurrent.futures import ThreadPoolExecutor
+from tqdm import tqdm
+
+
+class Evaluator:
+    def __init__(self):
+        self.parse_args()
+        self.init_metrics()
+        self.evaluate()
+        self.write_excel()
+        
+    def parse_args(self):
+        parser = argparse.ArgumentParser()
+        parser.add_argument('--pred-dir', type=str, required=True)
+        parser.add_argument('--true-dir', type=str, required=True)
+        parser.add_argument('--num-workers', type=int, default=48)
+        parser.add_argument('--metrics', type=str, nargs='+', default=[
+            'pha_mad', 'pha_mse', 'pha_grad', 'pha_dtssd', 'pha_conn'])
+        self.args = parser.parse_args()
+        
+    def init_metrics(self):
+        self.mad = MetricMAD()
+        self.mse = MetricMSE()
+        self.grad = MetricGRAD()
+        self.conn = MetricCONN()
+        self.dtssd = MetricDTSSD()
+        
+    def evaluate(self):
+        tasks = []
+        position = 0
+        
+        with ThreadPoolExecutor(max_workers=self.args.num_workers) as executor:
+            for dataset in sorted(os.listdir(self.args.pred_dir)):
+                if os.path.isdir(os.path.join(self.args.pred_dir, dataset)):
+                    for clip in sorted(os.listdir(os.path.join(self.args.pred_dir, dataset))):
+                        future = executor.submit(self.evaluate_worker, dataset, clip, position)
+                        tasks.append((dataset, clip, future))
+                        position += 1
+                    
+        self.results = [(dataset, clip, future.result()) for dataset, clip, future in tasks]
+        
+    def write_excel(self):
+        workbook = xlsxwriter.Workbook(os.path.join(self.args.pred_dir, f'{os.path.basename(self.args.pred_dir)}.xlsx'))
+        summarysheet = workbook.add_worksheet('summary')
+        metricsheets = [workbook.add_worksheet(metric) for metric in self.results[0][-1].keys()]
+        
+        for i, metric in enumerate(self.results[0][-1].keys()):
+            summarysheet.write(i, 0, metric)
+            summarysheet.write(i, 1, f'={metric}!B2')
+        
+        for row, (dataset, clip, metrics) in enumerate(self.results):
+            for metricsheet, metric in zip(metricsheets, metrics.values()):
+                # Write the header
+                if row == 0:
+                    metricsheet.write(1, 0, 'Average')
+                    metricsheet.write(1, 1, f'=AVERAGE(C2:ZZ2)')
+                    for col in range(len(metric)):
+                        metricsheet.write(0, col + 2, col)
+                        colname = xlsxwriter.utility.xl_col_to_name(col + 2)
+                        metricsheet.write(1, col + 2, f'=AVERAGE({colname}3:{colname}9999)')
+                        
+                metricsheet.write(row + 2, 0, dataset)
+                metricsheet.write(row + 2, 1, clip)
+                metricsheet.write_row(row + 2, 2, metric)
+        
+        workbook.close()
+
+    def evaluate_worker(self, dataset, clip, position):
+        framenames = sorted(os.listdir(os.path.join(self.args.pred_dir, dataset, clip, 'pha')))
+
+        metrics = {metric_name : [] for metric_name in self.args.metrics}
+        
+        pred_pha_tm1 = None
+        true_pha_tm1 = None
+        
+        for i, framename in enumerate(tqdm(framenames, desc=f'{dataset} {clip}', position=position, dynamic_ncols=True)):
+            true_pha = cv2.imread(os.path.join(self.args.true_dir, dataset, clip, 'pha', framename[:-4]+".png"), cv2.IMREAD_GRAYSCALE).astype(np.float32) / 255
+            pred_pha = cv2.imread(os.path.join(self.args.pred_dir, dataset, clip, 'pha', framename), cv2.IMREAD_GRAYSCALE).astype(np.float32) / 255
+            
+            if 'pha_mad' in self.args.metrics:
+                metrics['pha_mad'].append(self.mad(pred_pha, true_pha))
+            if 'pha_mse' in self.args.metrics:
+                metrics['pha_mse'].append(self.mse(pred_pha, true_pha))
+            if 'pha_grad' in self.args.metrics:
+                metrics['pha_grad'].append(self.grad(pred_pha, true_pha))
+            if 'pha_conn' in self.args.metrics:
+                metrics['pha_conn'].append(self.conn(pred_pha, true_pha))
+            if 'pha_dtssd' in self.args.metrics:
+                if i == 0:
+                    metrics['pha_dtssd'].append(0)
+                else:
+                    metrics['pha_dtssd'].append(self.dtssd(pred_pha, pred_pha_tm1, true_pha, true_pha_tm1))
+                    
+            pred_pha_tm1 = pred_pha
+            true_pha_tm1 = true_pha
+            
+
+        return metrics
+
+    
+class MetricMAD:
+    def __call__(self, pred, true):
+        return np.abs(pred - true).mean() * 1e3
+
+
+class MetricMSE:
+    def __call__(self, pred, true):
+        return ((pred - true) ** 2).mean() * 1e3
+
+
+class MetricGRAD:
+    def __init__(self, sigma=1.4):
+        self.filter_x, self.filter_y = self.gauss_filter(sigma)
+    
+    def __call__(self, pred, true):
+        pred_normed = np.zeros_like(pred)
+        true_normed = np.zeros_like(true)
+        cv2.normalize(pred, pred_normed, 1., 0., cv2.NORM_MINMAX)
+        cv2.normalize(true, true_normed, 1., 0., cv2.NORM_MINMAX)
+
+        true_grad = self.gauss_gradient(true_normed).astype(np.float32)
+        pred_grad = self.gauss_gradient(pred_normed).astype(np.float32)
+
+        grad_loss = ((true_grad - pred_grad) ** 2).sum()
+        return grad_loss / 1000
+    
+    def gauss_gradient(self, img):
+        img_filtered_x = cv2.filter2D(img, -1, self.filter_x, borderType=cv2.BORDER_REPLICATE)
+        img_filtered_y = cv2.filter2D(img, -1, self.filter_y, borderType=cv2.BORDER_REPLICATE)
+        return np.sqrt(img_filtered_x**2 + img_filtered_y**2)
+    
+    @staticmethod
+    def gauss_filter(sigma, epsilon=1e-2):
+        half_size = np.ceil(sigma * np.sqrt(-2 * np.log(np.sqrt(2 * np.pi) * sigma * epsilon)))
+        size = int(2 * half_size + 1)
+
+        # create filter in x axis
+        filter_x = np.zeros((size, size))
+        for i in range(size):
+            for j in range(size):
+                filter_x[i, j] = MetricGRAD.gaussian(i - half_size, sigma) * MetricGRAD.dgaussian(
+                    j - half_size, sigma)
+
+        # normalize filter
+        norm = np.sqrt((filter_x**2).sum())
+        filter_x = filter_x / norm
+        filter_y = np.transpose(filter_x)
+
+        return filter_x, filter_y
+        
+    @staticmethod
+    def gaussian(x, sigma):
+        return np.exp(-x**2 / (2 * sigma**2)) / (sigma * np.sqrt(2 * np.pi))
+    
+    @staticmethod
+    def dgaussian(x, sigma):
+        return -x * MetricGRAD.gaussian(x, sigma) / sigma**2
+
+
+class MetricCONN:
+    def __call__(self, pred, true):
+        step=0.1
+        thresh_steps = np.arange(0, 1 + step, step)
+        round_down_map = -np.ones_like(true)
+        for i in range(1, len(thresh_steps)):
+            true_thresh = true >= thresh_steps[i]
+            pred_thresh = pred >= thresh_steps[i]
+            intersection = (true_thresh & pred_thresh).astype(np.uint8)
+
+            # connected components
+            _, output, stats, _ = cv2.connectedComponentsWithStats(
+                intersection, connectivity=4)
+            # start from 1 in dim 0 to exclude background
+            size = stats[1:, -1]
+
+            # largest connected component of the intersection
+            omega = np.zeros_like(true)
+            if len(size) != 0:
+                max_id = np.argmax(size)
+                # plus one to include background
+                omega[output == max_id + 1] = 1
+
+            mask = (round_down_map == -1) & (omega == 0)
+            round_down_map[mask] = thresh_steps[i - 1]
+        round_down_map[round_down_map == -1] = 1
+
+        true_diff = true - round_down_map
+        pred_diff = pred - round_down_map
+        # only calculate difference larger than or equal to 0.15
+        true_phi = 1 - true_diff * (true_diff >= 0.15)
+        pred_phi = 1 - pred_diff * (pred_diff >= 0.15)
+
+        connectivity_error = np.sum(np.abs(true_phi - pred_phi))
+        return connectivity_error / 1000
+
+
+class MetricDTSSD:
+    def __call__(self, pred_t, pred_tm1, true_t, true_tm1):
+        dtSSD = ((pred_t - pred_tm1) - (true_t - true_tm1)) ** 2
+        dtSSD = np.sum(dtSSD) / true_t.size
+        dtSSD = np.sqrt(dtSSD)
+        return dtSSD * 1e2
+
+
+
+if __name__ == '__main__':
+    Evaluator()
\ No newline at end of file
diff --git a/evaluation/infer_batch_crgnn.sh b/evaluation/infer_batch_crgnn.sh
new file mode 100644
index 0000000..d5ce3b7
--- /dev/null
+++ b/evaluation/infer_batch_crgnn.sh
@@ -0,0 +1,31 @@
+#!/bin/bash
+
+input_folder="data/crgnn/image_allframe"
+mask_folder="data/crgnn/mask"
+
+ckpt_name="matanyone2"
+
+for video_folder in "${input_folder}"/*; do
+  if [ -d "${video_folder}" ]; then
+    video_id=$(basename "${video_folder}")
+
+    mask_file="${mask_folder}/${video_id}.png"
+    if [ -f "${mask_file}" ]; then
+
+      input_frames_folder="${video_folder}"
+      if [ -d "${input_frames_folder}" ]; then
+        echo "Processing video: ${video_id}"
+        
+        python inference_matanyone2.py \
+                --input_path "${input_frames_folder}" \
+                --mask_path "${mask_file}" \
+                --output_path "data/results/crgnn" \
+                --ckpt_path "pretrained_models/${ckpt_name}.pth" \
+                --warmup 10 \
+                --erode_kernel 10 \
+                --dilate_kernel 10 \
+                --save_image
+      fi
+    fi
+  fi
+done
diff --git a/evaluation/infer_batch_hr_vm.sh b/evaluation/infer_batch_hr_vm.sh
new file mode 100644
index 0000000..9d58e0a
--- /dev/null
+++ b/evaluation/infer_batch_hr_vm.sh
@@ -0,0 +1,38 @@
+#!/bin/bash
+
+input_folder="./data/VideoMatte/videomatte_1920x1080"
+mask_folder="./data/VideoMatte_first_frame_seg_mask/videomatte_1920x1080"
+
+ckpt_name="matanyone2"
+
+for subfolder in "videomatte_motion" "videomatte_static"; do
+  subfolder_path="${input_folder}/${subfolder}"
+  
+  echo "Processing subfolder: ${subfolder}"
+  
+  for video_folder in "${subfolder_path}"/*; do
+    if [ -d "${video_folder}" ]; then
+      video_id=$(basename "${video_folder}")
+
+      mask_file="${mask_folder}/${subfolder}/${video_id}.png"
+      if [ -f "${mask_file}" ]; then
+
+        input_frames_folder="${video_folder}/com"
+        if [ -d "${input_frames_folder}" ]; then
+          echo "Processing video: ${video_id} from ${subfolder}"
+          
+          python evaluation/inference_matanyone_eval.py \
+                  --input_path "${input_frames_folder}" \
+                  --mask_path "${mask_file}" \
+                  --output_path "./data/results/videomatte_1920x1080/${subfolder}" \
+                  --ckpt_path "pretrained_models/${ckpt_name}.pth" \
+                  --warmup 10 \
+                  --erode_kernel 15 \
+                  --dilate_kernel 15 \
+                  --save_image
+        fi
+      fi
+    fi
+  done
+done
+
diff --git a/evaluation/infer_batch_hr_yt.sh b/evaluation/infer_batch_hr_yt.sh
new file mode 100644
index 0000000..abf43b0
--- /dev/null
+++ b/evaluation/infer_batch_hr_yt.sh
@@ -0,0 +1,37 @@
+#!/bin/bash
+
+input_folder="./data/YouTubeMatte/youtubematte_1920x1080"
+mask_folder="./data/YouTubeMatte_first_frame_seg_mask/youtubematte_1920x1080"
+
+ckpt_name="matanyone2"
+
+for subfolder in "youtubematte_motion" "youtubematte_static"; do
+  subfolder_path="${input_folder}/${subfolder}"
+  
+  echo "Processing subfolder: ${subfolder}"
+  
+  for video_folder in "${subfolder_path}"/*; do
+    if [ -d "${video_folder}" ]; then
+      video_id=$(basename "${video_folder}")
+
+      mask_file="${mask_folder}/${video_id}.png"
+      if [ -f "${mask_file}" ]; then
+
+        input_frames_folder="${video_folder}/har"
+        if [ -d "${input_frames_folder}" ]; then
+          echo "Processing video: ${video_id} from ${subfolder}"
+          
+          python evaluation/inference_matanyone_eval.py \
+                  --input_path "${input_frames_folder}" \
+                  --mask_path "${mask_file}" \
+                  --output_path "./data/results/youtubematte_1920x1080/${subfolder}" \
+                  --ckpt_path "pretrained_models/${ckpt_name}.pth" \
+                  --warmup 10 \
+                  --erode_kernel 15 \
+                  --dilate_kernel 15 \
+                  --save_image
+        fi
+      fi
+    fi
+  done
+done
\ No newline at end of file
diff --git a/evaluation/infer_batch_lr_vm.sh b/evaluation/infer_batch_lr_vm.sh
new file mode 100644
index 0000000..806dba4
--- /dev/null
+++ b/evaluation/infer_batch_lr_vm.sh
@@ -0,0 +1,38 @@
+#!/bin/bash
+
+input_folder="./data/VideoMatte/videomatte_512x288"
+mask_folder="./data/VideoMatte_first_frame_seg_mask/videomatte_512x288"
+
+ckpt_name="matanyone2"
+
+for subfolder in "videomatte_motion" "videomatte_static"; do
+  subfolder_path="${input_folder}/${subfolder}"
+  
+  echo "Processing subfolder: ${subfolder}"
+  
+  for video_folder in "${subfolder_path}"/*; do
+    if [ -d "${video_folder}" ]; then
+      video_id=$(basename "${video_folder}")
+
+      mask_file="${mask_folder}/${subfolder}/${video_id}.png"
+      if [ -f "${mask_file}" ]; then
+
+        input_frames_folder="${video_folder}/com"
+        if [ -d "${input_frames_folder}" ]; then
+          echo "Processing video: ${video_id} from ${subfolder}"
+          
+          python evaluation/inference_matanyone_eval.py \
+                  --input_path "${input_frames_folder}" \
+                  --mask_path "${mask_file}" \
+                  --output_path "./data/results/videomatte_512x288/${subfolder}" \
+                  --ckpt_path "pretrained_models/${ckpt_name}.pth" \
+                  --warmup 1 \
+                  --erode_kernel 4 \
+                  --dilate_kernel 4 \
+                  --save_image
+        fi
+      fi
+    fi
+  done
+done
+
diff --git a/evaluation/infer_batch_lr_yt.sh b/evaluation/infer_batch_lr_yt.sh
new file mode 100644
index 0000000..65c52b9
--- /dev/null
+++ b/evaluation/infer_batch_lr_yt.sh
@@ -0,0 +1,37 @@
+#!/bin/bash
+
+input_folder="./data/YouTubeMatte/youtubematte_512x288"
+mask_folder="./data/YouTubeMatte_first_frame_seg_mask/youtubematte_512x288"
+
+ckpt_name="matanyone2"
+
+for subfolder in "youtubematte_motion" "youtubematte_static"; do
+  subfolder_path="${input_folder}/${subfolder}"
+  
+  echo "Processing subfolder: ${subfolder}"
+  
+  for video_folder in "${subfolder_path}"/*; do
+    if [ -d "${video_folder}" ]; then
+      video_id=$(basename "${video_folder}")
+
+      mask_file="${mask_folder}/${video_id}.png"
+      if [ -f "${mask_file}" ]; then
+
+        input_frames_folder="${video_folder}/har"
+        if [ -d "${input_frames_folder}" ]; then
+          echo "Processing video: ${video_id} from ${subfolder}"
+          
+          python evaluation/inference_matanyone_eval.py \
+                  --input_path "${input_frames_folder}" \
+                  --mask_path "${mask_file}" \
+                  --output_path "./data/results/youtubematte_512x288/${subfolder}" \
+                  --ckpt_path "pretrained_models/${ckpt_name}.pth" \
+                  --warmup 1 \
+                  --erode_kernel 4 \
+                  --dilate_kernel 4 \
+                  --save_image
+        fi
+      fi
+    fi
+  done
+done
\ No newline at end of file
diff --git a/evaluation/inference_matanyone_eval.py b/evaluation/inference_matanyone_eval.py
new file mode 100644
index 0000000..b115bac
--- /dev/null
+++ b/evaluation/inference_matanyone_eval.py
@@ -0,0 +1,154 @@
+import os
+import cv2
+import tqdm
+import numpy as np
+from PIL import Image
+
+import torch
+import torch.nn.functional as F
+
+from matanyone2.inference.inference_core import InferenceCore
+from matanyone2.utils.get_default_model import get_matanyone2_model
+from matanyone2.utils.device import get_default_device, safe_autocast_decorator
+from matanyone2.utils.inference_utils import gen_dilate, gen_erosion, read_frame_from_videos
+
+
+import warnings
+warnings.filterwarnings("ignore")
+
+device = get_default_device()
+
+@torch.inference_mode()
+@safe_autocast_decorator()
+def main(input_path, mask_path, output_path, ckpt_path, n_warmup=10, r_erode=10, r_dilate=10, suffix="", save_image=False, max_size=-1):
+
+    video_name = os.path.basename(os.path.dirname(input_path))
+
+    # load MatAnyone model
+    matanyone2 = get_matanyone2_model(ckpt_path, device)
+
+    # init inference processor
+    processor = InferenceCore(matanyone2, cfg=matanyone2.cfg)
+
+    # inference parameters
+    r_erode = int(r_erode)
+    r_dilate = int(r_dilate)
+    n_warmup = int(n_warmup)
+    max_size = int(max_size)
+
+    # load input frames
+    vframes, fps, length, _ = read_frame_from_videos(input_path)
+    repeated_frames = vframes[0].unsqueeze(0).repeat(n_warmup, 1, 1, 1) # repeat the first frame for warmup
+    vframes = torch.cat([repeated_frames, vframes], dim=0).float()
+    length += n_warmup  # update length
+
+    # resize if needed
+    if max_size > 0:
+        h, w = vframes.shape[-2:]
+        min_side = min(h, w)
+        if min_side > max_size:
+            new_h = int(h / min_side * max_size)
+            new_w = int(w / min_side * max_size)
+
+        vframes = F.interpolate(vframes, size=(new_h, new_w), mode="area")
+        
+    # set output paths
+    os.makedirs(output_path, exist_ok=True)
+    if suffix != "":
+        video_name = f'{video_name}_{suffix}'
+    if save_image:
+        os.makedirs(f'{output_path}/{video_name}', exist_ok=True)
+        os.makedirs(f'{output_path}/{video_name}/pha', exist_ok=True)
+        os.makedirs(f'{output_path}/{video_name}/fgr', exist_ok=True)
+
+    # load the first-frame mask
+    mask = Image.open(mask_path).convert('L')
+    mask = np.array(mask)
+
+    bgr = (np.array([120, 255, 155], dtype=np.float32)/255).reshape((1, 1, 3)) # green screen to paste fgr
+    objects = [1]
+
+    # [optional] erode & dilate
+    if r_dilate > 0:
+        mask = gen_dilate(mask, r_dilate, r_dilate)
+    if r_erode > 0:
+        mask = gen_erosion(mask, r_erode, r_erode)
+
+    mask = torch.from_numpy(mask).float().to(device)
+
+    if max_size > 0:  # resize needed
+        mask = F.interpolate(mask.unsqueeze(0).unsqueeze(0), size=(new_h, new_w), mode="nearest")
+        mask = mask[0,0]
+
+    # inference start
+    phas = []
+    fgrs = []
+    for ti in tqdm.tqdm(range(length)):
+        # load the image as RGB; normalization is done within the model
+        image = vframes[ti]
+
+        image_np = np.array(image.permute(1,2,0))       # for output visualize
+        image = (image / 255.).float().to(device)       # for network input
+
+        if ti == 0:
+            output_prob = processor.step(image, mask, objects=objects)      # encode given mask
+            output_prob = processor.step(image, first_frame_pred=True)      # first frame for prediction
+        else:
+            if ti <= n_warmup:
+                output_prob = processor.step(image, first_frame_pred=True)  # reinit as the first frame for prediction
+            else:
+                output_prob = processor.step(image)
+
+        # convert output probabilities to alpha matte
+        mask = processor.output_prob_to_mask(output_prob)
+
+        # visualize prediction
+        pha = mask.unsqueeze(2).cpu().numpy()
+        com_np = image_np / 255. * pha + bgr * (1 - pha)
+        
+        # DONOT save the warmup frame
+        if ti > (n_warmup-1):
+            com_np = np.round(np.clip(com_np * 255.0, 0, 255)).astype(np.uint8)
+            pha = np.round(np.clip(pha * 255.0, 0, 255)).astype(np.uint8)
+            fgrs.append(com_np)
+            phas.append(pha)
+            if save_image:
+                cv2.imwrite(f'{output_path}/{video_name}/pha/{str(ti-n_warmup).zfill(4)}.png', pha)
+                cv2.imwrite(f'{output_path}/{video_name}/fgr/{str(ti-n_warmup).zfill(4)}.png', com_np[...,[2,1,0]])
+
+    # [optional] save videos for better visualization
+    # import imageio
+    
+    # phas = np.array(phas)
+    # fgrs = np.array(fgrs)
+
+    # imageio.mimwrite(f'{output_path}/{video_name}_fgr.mp4', fgrs, fps=fps, quality=7)
+    # imageio.mimwrite(f'{output_path}/{video_name}_pha.mp4', phas, fps=fps, quality=7)
+
+if __name__ == '__main__':
+    import argparse
+    parser = argparse.ArgumentParser()
+    parser.add_argument('-i', '--input_path', type=str, default="inputs/video/test-sample1.mp4", help='Path of the input video or frame folder.')
+    parser.add_argument('-m', '--mask_path', type=str, default="inputs/mask/test-sample1.png", help='Path of the first-frame segmentation mask.')
+    parser.add_argument('-o', '--output_path', type=str, default="results/", help='Output folder. Default: results')
+    parser.add_argument('-c', '--ckpt_path', type=str, default="pretrained_models/matanyone.pth", help='Path of the MatAnyone model.')
+    parser.add_argument('-w', '--warmup', type=str, default="10", help='Number of warmup iterations for the first frame alpha prediction.')
+    parser.add_argument('-e', '--erode_kernel', type=str, default="10", help='Erosion kernel on the input mask.')
+    parser.add_argument('-d', '--dilate_kernel', type=str, default="10", help='Dilation kernel on the input mask.')
+    parser.add_argument('--suffix', type=str, default="", help='Suffix to specify different target when saving, e.g., target1.')
+    parser.add_argument('--save_image', action='store_true', default=False, help='Save output frames. Default: False')
+    parser.add_argument('--max_size', type=str, default="-1", help='When positive, the video will be downsampled if min(w, h) exceeds. Default: -1 (means no limit)')
+
+    
+    args = parser.parse_args()
+
+    main(input_path=args.input_path, \
+         mask_path=args.mask_path, \
+         output_path=args.output_path, \
+         ckpt_path=args.ckpt_path, \
+         n_warmup=args.warmup, \
+         r_erode=args.erode_kernel, \
+         r_dilate=args.dilate_kernel, \
+         suffix=args.suffix, \
+         save_image=args.save_image, \
+         max_size=args.max_size)