MODNet

MODNet – Fast Image Matting for Background Removal

This repository provides a Hugging Face–compatible version of MODNet (Mobile Object Detection Network) for fast and high-quality foreground matting and background removal.

The model is designed to produce pixel-perfect alpha mattes, handling fine details such as hair, fabric edges, and soft shadows, while remaining efficient enough for real-time CPU inference.

---

🔹 Model Details

• Architecture: MODNet (custom architecture)
• Task: Image matting / background removal
• Framework: PyTorch
• Backbone: MobileNetV2
• Input: RGB image tensor (B, 3, H, W)
• Output: (semantic, detail, matte) predictions

How to Load the Model

from transformers import AutoModel

model = AutoModel.from_pretrained(
    "boopathiraj/MODNet",
    trust_remote_code=True
)

model.eval()

Example Inference

import torch
import cv2
import numpy as np
from PIL import Image
from torchvision import transforms

# Preprocess
def preprocess(image_path, ref_size=512):
    image = Image.open(image_path).convert("RGB")
    w, h = image.size

    # Resize while maintaining aspect ratio
    scale = ref_size / max(h, w)
    new_w, new_h = int(w * scale), int(h * scale)
    image_resized = image.resize((new_w, new_h), Image.BILINEAR)

    # Create a new blank image of ref_size x ref_size and paste the resized image
    # This will pad the image to ref_size x ref_size
    padded_image = Image.new("RGB", (ref_size, ref_size), (0, 0, 0)) # Black padding
    padded_image.paste(image_resized, ((ref_size - new_w) // 2, (ref_size - new_h) // 2))

    transform = transforms.Compose([
        transforms.ToTensor(),
        transforms.Normalize(mean=[0.5, 0.5, 0.5],
                             std=[0.5, 0.5, 0.5])
    ])

    tensor = transform(padded_image).unsqueeze(0)
    return tensor, (w, h)

# Run inference
inp, original_size = preprocess("/content/portrait.jpg")

import torch
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")

model.to(device)
inp = inp.to(device) # Assign the tensor to the device

with torch.no_grad():
    semantic, detail, matte = model(inp, True)

matte = matte[0, 0].cpu().numpy()
matte = cv2.resize(matte, original_size)

# Save alpha matte
matte = (matte * 255).astype(np.uint8)
Image.fromarray(matte).save("alpha_matte.png")

Visualization

import numpy as np
from PIL import Image

def combined_display(image, matte):
  # calculate display resolution
  w, h = image.width, image.height
  rw, rh = 800, int(h * 800 / (3 * w))
  
  # obtain predicted foreground
  image = np.asarray(image)
  if len(image.shape) == 2:
    image = image[:, :, None]
  if image.shape[2] == 1:
    image = np.repeat(image, 3, axis=2)
  elif image.shape[2] == 4:
    image = image[:, :, 0:3]
  matte = np.repeat(np.asarray(matte)[:, :, None], 3, axis=2) / 255
  foreground = image * matte + np.full(image.shape, 255) * (1 - matte)
  
  # combine image, foreground, and alpha into one line
  combined = np.concatenate((image, foreground, matte * 255), axis=1)
  combined = Image.fromarray(np.uint8(combined)).resize((rw, rh))
  return combined

# visualize all images
image_names = os.listdir(input_folder)
for image_name in image_names:
  matte_name = image_name.split('.')[0] + '.png'
  image = Image.open(os.path.join(input_folder, image_name))
  matte = Image.open(os.path.join(output_folder, matte_name))
  display(combined_display(image, matte))
  print(image_name, '\n')

---