Anime Super Resolution 🖼️

Introduction 🌟

Anime Super Resolution 🖼️ enhances anime-style images using a fine-tuned Real-ESRGAN model, optimized for clarity and detail. Built on the RealESRGAN_x4plus model, it leverages a private dataset of 27,052 high-resolution (1920x1080) anime frames 📸. The project offers tools for data processing, training, inference, and an interactive Gradio demo, accessible on platforms like Colab, Kaggle, and locally 🚀.

Key Features ✨

• Anime-Specific Upscaling 🎨: Fine-tuned Real-ESRGAN for high-quality anime image super-resolution.

• Large Anime Dataset 📚: 27,052 high-res anime frames for robust training.

• Interactive Gradio Demo 🖥️: Easy model testing via HuggingFace-hosted interface.

• Multi-Platform Support 🌐: Runs on Colab, Kaggle, JupyterLab, and more.

• Data Processing Tools 🛠️: Scripts for multiscale dataset creation and meta-info generation.

• Flexible Training/Inference ⚙️: Customizable configurations for training and upscaling.

• Open-Source 📖: MIT-licensed, built with PyTorch, NumPy, and Pillow.

• Local/Cloud Compatibility ☁️: Supports local Gradio app and cloud-based execution.

Notebook

This notebook provides a step-by-step guide to finetune the Real-ESRGAN model for enhancing anime-style images. It covers data preparation, model configuration, training, and evaluation, optimized for clarity and reproducibility.

Dataset

Source

The dataset is privately extracted from anime films, comprising 27,052 images with a resolution of 1920x1080. Please read at Download Dataset if you want to use this dataset.

Data Structure

The Data Structure is

data/ 📁
├── anime-images-raw/ 📁
│   ├── frame_0001.jpg 📸
│   ├── frame_0001_1.jpg 📷
│   └── ... 📸
├── anime-images-multiscale/ 📁
│   ├── frame_0001T0.png 📸
│   ├── frame_0001T1.png 📸
│   ├── frame_0001T2.png 📸
│   ├── frame_0001T3.png 📸
│   ├── frame_0001_10T0.png 📸
│   ├── frame_0001_10T1.png 📸
│   ├── frame_0001_10T2.png 📸
│   ├── frame_0001_10T3.png 📸
│   └── ... 📸
└── meta_info/ 📁
    └── meta_info_multiscale.txt 📄

To proceed with the topic, please consult the section on Real-ESRGAN Data Processing for Training for comprehensive details and guidance.

Base Model

The Real-ESRGAN-Anime-finetuning model was developed by fine-tuning the pre-trained model, leveraging its robust foundation to enhance performance specifically for anime-style image super-resolution tasks.

Demonstration

Interactive Demo

Explore the interactive demo hosted on HuggingFace:

Below is a screenshot of the SlimFace Demo GUI:

Run Locally

To run the Gradio application locally at the default address localhost:7860, execute:

python apps/gradio_app.py

Installation

Clone GitHub Repository

git clone https://github.com/danhtran2mind/Anime-Super-Resolution
cd Anime-Super-Resolution

Install Dependencies (Training + Inference)

pip install -e .

Install Dependencies for Inference only

pip install -r requirements/requirements_inference.txt

Script-Driven Setup

Download Model checkpoints

• Download only model checkpoint for Inference

  python scripts/download_ckpts.py

• For training from Real-ESRGAN Model checkpoint

  python scripts/download_ckpts.py --base_model_only

• For training from Real-ESRGAN-Anime-finetuning Model checkpoint

  python scripts/download_ckpts.py --full_ckpts

More detail you can read at Download Model Checkpoints.

Setup Third Party

    python scripts/setup_third_party.py

Download Dataset

    python scripts/download_datasets.py \
        --dataset_id "<huggingface_dataset_id>"
        --huggingface_token "<your_huggingface_token>"

More detail you can read at Download Dataset.

Usage

Real-ESRGAN Data Processing (for Training)

First you need see the data structure at Data Structure. Then execute below scripts to process dataset.

• Create multiscale Folder

  python src/third_party/Real-ESRGAN/scripts/generate_multiscale_DF2K.py \
      --input ./data/anime-images-raw \
      --output ./data/anime-images-multiscale

• Create meta_info_multiscale.txt

  python src/third_party/Real-ESRGAN/scripts/generate_meta_info.py \
  --input ./data/anime-images-raw ./data/anime-images-multiscale \
  --root ./data ./data \
  --meta_info "./data/meta_info/meta_info_multiscale.txt"

Training

Training Script

python src/anime_super_resolution/train.py \
    --config "<your_model_config_yml_path>" \
    --auto_resume

Additional Arguments

For more details and available options, refer to the Training Document.

You can see Real-ESRGAN Training, and BasicSR Training Options for more details.

Inference

Inference Script

python src/anime_super_resolution/infer.py \
    --input_path tests/test_data/input_image.png \
    --output_dir tests/test_data \
    --suffix real_esrgan_anime \
    --outscale 2 \
    --model_path ckpts/Real-ESRGAN-Anime-finetuning/net_g_latest.pth

Inference Examples

Example	Image Type	Image
Ex. 1	Original Image
	Upscaled Image (2x)
Ex. 2	Original Image
	Upscaled Image (4x)
Ex. 3	Original Image
	Upscaled Image (6x)
Ex. 4	Original Image
	Upscaled Image (8x)

Additional Arguments

For more details and available options, refer to the Inference Document.

Environment

SlimFace requires the following environment:

• Python: 3.10 or higher
• Key Libraries: Refer to Requirements Compatible for compatible dependencies.

Project Credits and Resources

• This project leverages code from:

  The Original Real-ESRGAN by at . Our own bug fixes and enhancements are available at .

  The Inference code by at . Our own bug fixes and enhancements are available at

• You can explore more Model Hubs at:

  HuggingFace Model Hub: . Real-ESRGAN Model releases: . You also download Real-ESRGAN-Anime-finetuning at