tammy Download - tammy Source code download

tammy

Other source code

1.0.0

Download

Tammy

Tammy is a Python/Pytorch-based open-source project that uses deep learning models to generate original music videos. It allows users to automatically generate videos based on text prompt transitions that are synchronized with various aspects of a song, such as its BPM or piano pattern. The project utilizes deep learning models at various stages of the video generation process, including audio source separation with LSTMs, frame generation with GANs, spatial upscaling with super-resolution models, and temporal upsampling with frame interpolation models. The aim of this project is to provide an easy-to-use framework to build custom model pipelines to create unique music videos.

Example (turn on audio!)

video_with_audio.mp4

Features

fully automated music-video generation. Provide just a song, text prompts and generation settings to generate a music-video.
multiple animation modes: Animation_2d and Interpolation.
automatic and lazy model loading. For example, the Stable Diffusion model is automatically fetched when used so no need to manually install any libraries or wait for all models to be downloaded when installing the package.

Quick start

For a quick start:

Make sure to install ffmpeg and libsndfile sudo apt-get install ffmpeg libsndfile1 and git-lfs.
Clone this repo
Install tammy by running pip install .
run python run_tammy.py which will use the default settings in settingssettings_cpu.yaml and default song thoughtsarebeings_clip.wav.

The tammy package can be easily used in your own script or other setting files and audio files can be used with the existing run_tammy.py script by running python run_tammy.py --settings_file --audio_path .

Dataflow and Code Structure

tammy.prompthandler generates the settings for every frame to be generated (e.g. translation or text prompt) based on a more concise description of the generation settings.
tammy.sequence_maker has a generator which generates an image sequence based on a text prompts. Currently the supported models are VQGAN-CLIP and Stable-Diffusion
tammy.upscaling scales up the generated images with super-resolution. Currently the only supported model is SwinIR.
tammy.superslowmo interpolates generated (optionally upscaled) images to increase the FPS without needing to generate every frame with a sequence_maker. Currently the only supported model is SuperSloMo.

Generation Settings

The video generation has many configuration settings which are specified in a .yaml file. Some example setting files, mostly used for testing, can be found in the settings folder. Most setting names (keys in the settings.yaml) should be self-explanatory. For clarity, some settings are explained below.

Instrument

Instruments are used to steer frame transitions, in particular: zoom in Animation_2d mode and prompt transition speed in Interpolation mode. tammy has two options to provide instruments:

automatically by using Spleeter source separation: set do_spleet: True and provide instrument:
manually by providing a keyframe file in the setting file as zoom_instrument: and name the file: file_name_fps.txt where fps should correspond with the fps value in sequence_settings.initial_fps . Keyframes can be manually generated with e.g. https://www.chigozie.co.uk/audio-keyframe-generator/

Frame rate and video length

The setting sequence_settings.initial_fps determines the number of frames generated, given the length of the audio clip. By using frame interpolation, the frame-rate can be increased to a target by setting do_slowmo: True and providing a target_fps which must be a multiple of initial_fps. This allows to produce high frame rate videos faster than compared to generating all frames from scratch with the generator.

If desired, the number of generated frames can be limited by providing sequence_settings.max_frames. In this case the generated video length will be shorter than the provided audio clip and will be: max_frames/initial_fps.