We explore the creation of a dynamic bridge between between two paintings, seamlessly transitioning between artworks adding motion to static images. We build a two-stage generative process, creating both abstract-conceptual interpolation as well as spatio-temporal interpolation. We first hallucinate intermediate artworks using a generative diffusion model, then interpolate between resulting frames using a motion generation network, thus obtaining a complete video out of static paintings.
final_video.mp4
final_video.mp4
Our proposed method consists of a three-stage generative process.
-
Conceptual interpolation stage: we employ a diffusion-based image generative model Razzhigaev et al. 2023 to efficiently produce π intermediate images πΎ1 to πΎπ situated between two reference paintings, A and B.
-
Spatio-temporal interpolation stage: we employ a large motion interpolation network Reda et al. 2022 to generate motion sequences between A, the generated images πΎ and B.
-
Upscaling stage: we finally upscale our results by employing a video super-resolution model Wang et al. reaching a resolution of 2K.
You can see the full pipeline in the following illustration.
Install ffmpeg and av dev libs
sudo apt install ffmpeg libavformat-dev libavdevice-dev
- Diffusion environment
conda create -p /projects/Anaconda/envs/diff python=3.8 -y
conda activate diff
conda install cudnn -y
pip install -r requirements_total.txt --no-cache-dir
pip install -U git+https://github.com/facebookresearch/demucs#egg=demucs
export LD_LIBRARY_PATH=$LD_LIBRARY_PATH:/projects/Anaconda/envs/diff/lib/python3.8/site-packages/tensorrt_libs
pip install "git+https://github.com/ai-forever/Kandinsky-2.git"
pip install git+https://github.com/openai/CLIP.git
conda deactivate
- Create Image Interpolation environemnt
conda create -n film python=3.9 -y
conda activate film
cd frame-interpolation
pip install -r requirements.txt
pip install tensorflow
conda deactivate
- (Optional) Create Music environment
conda create -n music python=3.8
conda activate srmusic
pip install torch torchvideo torchaudio
pip install -U git+https://github.com/facebookresearch/demucs#egg=demucs
pip install tqdm
pip install matplotlib opencv-python librosa Ipython
conda deactivate
In order to use ArtWalks you just need to make a folder with all your favourite images. By launching the following script you will be able to generate a 2K video with your own pictures:
bash full_pipeline.sh $INPATH $FOLDER_NAME $S $I $F $MODEL
Where:
- INPATH : Path to the input data folder
- FOLDER_NAME : Name of the folder containing the images
- S : Number of images to generate with diffusion between each pair of images
- I : Number of seconds that have to pass between each consecutive pair of generated images
- F : Number of seconds to freeze on each original image during the video
- MODEL : Name of the diffusion model to use (unclip/kandinsky)
When you have a big collection of images, you can use this tool to select a random subset of img_num images. Generate a random story from a collection of immages:
python random_story_generation.py --input_path inputdata --img_num 10
Where:
- input_path : Path to the folder containing all subfolders of images
- img_num : Number of images to randomly sample
You can also just generate any sequence of images between two given pictures.
Run diffusion pipeline to interpolate between every pair of images in input_path
(It takes roughly ~10s per pair):
python diffusion_models/diffusion.py --input_path $INPATH --folder_name $FOLDER_NAME --output_path $OUTPATH --model $M --interpolation_steps $S --outpaint
This script makes use of either the UnCLIP Image Interpolation pipeline or the Kandisnky 2 model. Script arguments:
parser.add_argument("--input_path", help="Path to folder with images",
type=str)
parser.add_argument("--folder_name", help="Name of the folder to read",
type=str)
parser.add_argument("--output_path", help="Path to the output folder",
type=str, default="output")
parser.add_argument("--model", help="Choose between kandinsky/unclip model", type=str, default='unclip')
parser.add_argument("--interpolation_steps", help="Number of generated frames between a pair of images",
type=int, default=5)
parser.add_argument("--square_crop", help="If active, crops the images in a square.", action="store_true")
parser.add_argument("--no_originals", help="If active, don't save original images.", action="store_true")
parser.add_argument("--no_originals", help="If active, don't save original images.", action="store_true")
parser.add_argument("--outpaint", help="If active, outpaints the image using generative model.", action="store_true")
parser.add_argument("--h", help="Height of the generated images",type=int, default=720)
parser.add_argument("--w", help="Width of the generated images",type=int, default=1280)
Generate videos interpolating between diffusion frames with (It takes roughly ~2 mins per pair):
python generate_videos_alpha.py --input_path output --folder_name $FOLDER_NAME --sec_interpolation $I --sec_freeze $F --clean
generate_videos_alpha.py
will produce videos with a non-linear acceleration between frames.
generate_videos.py
will produce videos with a linear interpolation speed between frames.
Script arguments:
parser.add_argument("--input_path", help="Path to folder with images", default='output',
type=str)
parser.add_argument("--folder_name", help="Name of the folder to read",
type=str)
parser.add_argument("--sec_interpolation", help="Number of seconds to interpolate between images", type=int, default=10)
parser.add_argument("--sec_freeze", help="Number of seconds to freeze per original image", type=int, default=20)
parser.add_argument("--clean", help="Delete everything but the final video", action='store_true')
Generate up to 4K video by using Video Super Resolution ESRGAN
python inference_realesrgan_video.py -n RealESRGAN_x4plus -i ../output/$FOLDER_NAME/final_video.mp4 -o ../output/$FOLDER_NAME/ -s $sr
Script arguments:
parser.add_argument('-i', '--input', type=str, default='inputs', help='Input video, image or folder')
parser.add_argument(
'-n',
'--model_name',
type=str,
default='realesr-animevideov3',
help=('Model names: realesr-animevideov3 | RealESRGAN_x4plus_anime_6B | RealESRGAN_x4plus | RealESRNet_x4plus |'
' RealESRGAN_x2plus | realesr-general-x4v3'
'Default:realesr-animevideov3'))
parser.add_argument('-o', '--output', type=str, default='results', help='Output folder')
parser.add_argument(
'-dn',
'--denoise_strength',
type=float,
default=0.5,
help=('Denoise strength. 0 for weak denoise (keep noise), 1 for strong denoise ability. '
'Only used for the realesr-general-x4v3 model'))
parser.add_argument('-s', '--outscale', type=float, default=4, help='The final upsampling scale of the image')
parser.add_argument('--suffix', type=str, default='out', help='Suffix of the restored video')
parser.add_argument('-t', '--tile', type=int, default=0, help='Tile size, 0 for no tile during testing')
parser.add_argument('--tile_pad', type=int, default=10, help='Tile padding')
parser.add_argument('--pre_pad', type=int, default=0, help='Pre padding size at each border')
parser.add_argument('--face_enhance', action='store_true', help='Use GFPGAN to enhance face')
parser.add_argument(
'--fp32', action='store_true', help='Use fp32 precision during inference. Default: fp16 (half precision).')
parser.add_argument('--fps', type=float, default=None, help='FPS of the output video')
parser.add_argument('--ffmpeg_bin', type=str, default='ffmpeg', help='The path to ffmpeg')
parser.add_argument('--extract_frame_first', action='store_true')
parser.add_argument('--num_process_per_gpu', type=int, default=1)
parser.add_argument(
'--alpha_upsampler',
type=str,
default='realesrgan',
help='The upsampler for the alpha channels. Options: realesrgan | bicubic')
parser.add_argument(
'--ext',
type=str,
default='auto',
help='Image extension. Options: auto | jpg | png, auto means using the same extension as inputs')