MI-GAN

MI-GAN: A Simple Baseline for Image Inpainting on Mobile Devices

This repository is the official implementation of MI-GAN.

MI-GAN: A Simple Baseline for Image Inpainting on Mobile Devices
Andranik Sargsyan, Shant Navasardyan, Xingqian Xu, Humphrey Shi

[Paper] | [Supplementary]

Our method MI-GAN can produce plausible results both on complex scene images as well as on face images. The bubble chart on the right shows the advantage of our network over state-of-the-art approaches. The size of the bubble signifies the relative number of parameters of each approach and the number inside of the bubble shows the absolute number of model parameters. Our approach achieves a low FID, while being one order of magnitude smaller and faster than recent SOTA approaches.

Third-Party Implementations

(Feel free to share your app/implementation/demo by creating an issue)

• 🔥 https://inpaintweb.lxfater.com - an open-source tool for in-browser inpainting by @lxfater, [GitHub project]
• 📱 inpaint_wechat - a WeChat mini-program for on-device inpainting by @zhiyuan, [GitHub project]
• 💻 IOPaint - a free and open-source inpainting tool powered by SOTA AI models by @Sanster, [GitHub project]
  • To use MI-GAN model with IOPaint run

    pip install iopaint
    iopaint start --model migan
    

    in the terminal and open http://localhost:8080/

See For Developers section below for easy-to-use MI-GAN ONNX Pipeline creation script, and for a pre-generated ONNX file, which can be seamlessly integrated into your app.

Prepare environment

Prepare the environment (Python>=3.8 is needed):

sudo apt install python3-venv
python3 -m venv venv
source ./venv/bin/activate
pip install pip --upgrade
pip install -r requirements.txt

Download pre-trained MI-GAN models from here (see LICENSE-WEIGHTS for the license) and put into ./models directory. If you also want to test with Co-Mod-GAN models, download pre-trained models from here and put into ./models directory.

Quick Test

For getting visual results, use demo.py script like this

python -m scripts.demo \
    --model-name migan-512 \
    --model-path ./models/migan_512_places2.pt \
    --images-dir ./examples/places2_512_object/images \
    --masks-dir ./examples/places2_512_object/masks \
    --output-dir ./examples/places2_512_object/results/migan \
    --device cuda \
    --invert-mask

For running with Co-Mod-GAN, change --model-name and --model-path accordingly. Also, depending on what the white color means in your masks, you may or may not need --invert-mask flag. By default white color denotes the known region.

To run inference on provided FFHQ samples, run

python -m scripts.demo \
    --model-name migan-256 \
    --model-path ./models/migan_256_ffhq.pt \
    --images-dir ./examples/ffhq_256_freeform/images \
    --masks-dir ./examples/ffhq_256_freeform/masks \
    --output-dir ./examples/ffhq_256_freeform/results/migan \
    --device cuda

You can find more example images and masks in examples directory, and modify the above command to reproduce all MI-GAN and Co-Mod-GAN results presented in the Supplementary material.

Note: With our provided free-form masks, --invert-mask option should not be used with the command.

Note: Even though our method is fully convolutional, the provided lib/model_zoo/migan_inference.py implementation currently performs some fixed-resolution operations. If you need fully convolutional support, you need to make filter_const and noise_const computations dynamic (depending on the input size). See self.register_buffer in SeparableConv2d class and in Upsample2d class.

Evaluation

For MI-GAN evaluation on Places2 download Places365-Standard 256x256 and high resolution validation images from Places2 official website.

Then, for evaluating MI-GAN FID and LPIPS on Places2 256x256 run

python -m scripts.evaluate_fid_lpips \
    --model-name migan-256 \
    --model-path ./models/migan_256_places2.pt \
    --real-dir PATH_TO_PLACES256_VALIDATION_IMAGES

For evaluating MI-GAN FID and LPIPS on Places2 512x512 run

python -m scripts.evaluate_fid_lpips \
    --model-name migan-512 \
    --model-path ./models/migan_512_places2.pt \
    --real-dir PATH_TO_PLACES512_VALIDATION_IMAGES

For MI-GAN evaluation on FFHQ 256x256 download FFHQ dataset from the official source, copy the first 10,000 samples into a separate validation directory, and then run

python -m scripts.evaluate_fid_lpips \
    --model-name migan-256 \
    --model-path ./models/migan_256_ffhq.pt \
    --real-dir PATH_TO_FFHQ_VALIDATION_IMAGES

The above commands will generate new free-form masks on the fly for doing the inference, but you can also download our pre-generated masks for Places2 256x256, Places2 512x512 and FFHQ 256x256 validation sets from here and use those for metrics calculations by providing an additional --mask-dir argument in above commands, where --mask-dir will point to the respective pre-generated mask directory.

Besides, you can generate your own free form masks using our script like so

python -m scripts.generate_masks \
    --images-dir PATH_TO_ORIGINAL_IMAGES_DIR \
    --output-dir MASK_OUTPUT_PATH \
    --resolution MASK_RESOLUTION

For Co-Mod-GAN metrics calculation change --model-name and --model-path accordingly. For example, in order to calculate Co-Mod-GAN metrics on FFHQ 256x256, run

python -m scripts.evaluate_fid_lpips \
    --model-name comodgan-256 \
    --model-path ./models/comodgan_256_ffhq.pt \
    --real-dir PATH_TO_FFHQ_VALIDATION_IMAGES

If your GPU doesn’t have enough memory for running the evaluation with default batch-size, you can reduce batch size by providing --batch-size argument with a low batch size in above commands.

FLOPs and parameter count calculation

In order to calculate FLOPs and parameter counts for MI-GAN and Co-Mod-GAN models you can run

python -m scripts.calculate_flops

Training

In order to smoothly run the training without changes, you need to have the following directory structure for datasets (data folder should be in the root project folder).

data
 | 
 +-- ffhq
 |  |  
 |  +-- ffhq256x256.zip
 |    
 +-- Places2
 |  |  
 |  +-- train_256
 |  +-- val_256
 |  +-- train_512
 |  +-- val_512

Also you need to download pre-trained Co-Mod-GAN models from here and put into models directory. models directory should have the following structure:

models
|  
| +-- comodgan_256_ffhq.pt
| +-- comodgan_256_places2.pt
| +-- comodgan_512_places2.pt

Then, for training MI-GAN 256 on Places 2 using 8 GPUs, run

./run.sh migan_places256 gpu01234567

For training MI-GAN 512 on Places 2 using 8 GPUs, run

./run.sh migan_places512 gpu01234567

For training MI-GAN 256 on FFHQ using 8 GPUs, run

./run.sh migan_ffhq256 gpu01234567

You can choose to run on specific GPUs. For example, if you want to run on GPUs 0,1,2,3, run the above commands with argument gpu0123. For more details, please see run.sh helper script.

After training you can generate the inference model from the saved best checkpoint with this command:

python -m scripts.export_inference_model \
    --model-path PATH_TO_BEST_PLACES512_PKL \
    --origs-dir ./examples/places2_512_freeform/images \
    --masks-dir ./examples/places2_512_freeform/masks \
    --output-dir ./exported_models/migan_places512 \
    --resolution 512 

The above example is for places2-512 model. You need to change the --resolution argument to 256 for 256 models. The exported onnx and pt models and sample results will be saved in the specified --output-dir directory.

Co-Mod-GAN training

If you want to conduct a training on your own dataset, you need to train the teacher network Co-Mod-Gan on your specific dataset before proceeding to MI-GAN training.

Here we provide example of how to train Co-Mod-GAN on Places and FFHQ datasets which were used in our experiments.

For training Co-Mod-GAN 256 on Places 2 using 8 GPUs, run

./run.sh comodgan_places256 gpu01234567

For training Co-Mod-GAN 512 on Places 2 using 8 GPUs, run

./run.sh comodgan_places512 gpu01234567

For training Co-Mod-GAN 256 on FFHQ using 8 GPUs, run

./run.sh comodgan_ffhq256 gpu01234567

You can choose to run on specific GPUs as in the case of MI-GAN training.

After the Co-Mod-GAN training is finished, you should modify teacher1_path in the corresponding configs/experiment/migan_*.yaml config to point to the respective Co-Mod-GAN’s best checkpoint and start training the MI-GAN as described in above section.

For Developers

We provide a script to convert the whole MI-GAN pipeline into ONNX for easier use in real-world applications. Please see scripts/create_onnx_pipeline.py for details. Here is an example usage of the script:

python -m scripts.create_onnx_pipeline \
    --resolution 512 \
    --model-path ./models/migan_512_places2.pt \
    --images-dir ./examples/places2_512_object/images \
    --masks-dir ./examples/places2_512_object/masks \
    --output-dir ./exported_models/places2_512_onnx_pipeline \
    --device cpu \
    --invert-mask

The generated ONNX model expects image and mask as an input, where image is a uint8 RGB image, and mask is a uint8 Grayscale mask (where 255 denotes known region values, 0 denotes masked region).

Pipeline does almost all necessary preprocessing (uint8->float32 conversion, cropping around the mask, resize to 512x512, normalization) and postprocessing (resizing to original, blending, float32->uint8 conversion). You just need to provide the right image and mask (must be binary).

The Pipeline supports arbitrary resolution inputs, but please note that when inpainting high resolution images the best results can be achieved with small, incremental brush strokes (if the inpainting area is large).

Pre-converted ONNX file for MI-GAN-512-Places2 Pipeline can be found here.

Note for researchers: The inference times reported in the paper are not based on the provided ONNX Pipeline. We measured the pure model inference times, whereas ONNX Pipeline introduced in this section includes complex pre-processing and post-processing, which are not included in reported times.

BibTeX

If you use our work in your research, please cite our publication:

@InProceedings{Sargsyan_2023_ICCV,
    author    = {Sargsyan, Andranik and Navasardyan, Shant and Xu, Xingqian and Shi, Humphrey},
    title     = {MI-GAN: A Simple Baseline for Image Inpainting on Mobile Devices},
    booktitle = {Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV)},
    month     = {October},
    year      = {2023},
    pages     = {7335-7345}
}

Acknowledgement

Our code is built upon SH-GAN.