fast-neural-style-transfer

Fast Neural Style Transfer (Picasso Style)

Apply a Picasso art style to any content image in under a second. Try it live in the browser →

This is a PyTorch implementation of "Perceptual Losses for Real-Time Style Transfer and Super-Resolution" (Johnson et al., 2016), trained as a feed-forward network for a single fixed style. Inference takes ~1 second on CPU and under a second on GPU.

Pushed to the Hub using the PytorchModelHubMixin integration.

• Code: github.com/HajarHAMDOUCH01/Fast-neural-style-transfer
• Paper: arXiv:1603.08155
• Demo: real-time-nst-app.streamlit.app

Usage

import torch
from huggingface_hub import hf_hub_download
import sys, os
import torchvision.transforms as transforms
from PIL import Image

# Download model file
model_file = hf_hub_download(
    repo_id="hajar001/fast-neural-style-transfer",
    filename="style_transfer_model.py"
)
sys.path.insert(0, os.path.dirname(model_file))
from style_transfer_model import StyleTransferNet

transform = transforms.Compose([
    transforms.Resize((256, 256)),
    transforms.ToTensor(),
])

# Load model
model = StyleTransferNet.from_pretrained("hajar001/fast-neural-style-transfer")
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
model = model.to(device)
model.eval()

with torch.no_grad():
    # Load and preprocess content image
    test_image = Image.open("content_image.jpg").convert("RGB")
    test_tensor = transform(test_image).unsqueeze(0).to(device)

    # Generate stylized image
    stylized_tensor = model(test_tensor)

    # Denormalize and convert to PIL
    denorm = transforms.Normalize(
        mean=[-0.485/0.229, -0.456/0.224, -0.406/0.225],
        std=[1/0.229, 1/0.224, 1/0.225]
    )
    stylized_tensor = denorm(stylized_tensor[0])
    stylized_tensor = torch.clamp(stylized_tensor, 0, 1)

    stylized_img = transforms.ToPILImage()(stylized_tensor.cpu())
    stylized_img.save("stylized_image.jpg")

Model Details

• Architecture: Feed-forward CNN with residual blocks, Instance Normalization, trained with perceptual + total variation loss (see training code for full details)
• Input: RGB image, resized to 256×256
• Output: Stylized RGB image, same resolution as input
• Style: Fixed at training time (Picasso) — this model applies one style; it is not a universal/arbitrary style transfer model