Open Source Windows AI Video Generators

Text2Video is a software tool that converts text to video for more engaging learning experience. I started this project because during this semester, I have been given many reading assignments and I felt frustration in reading long text. For me, it was very time and energy-consuming to learn something through reading. So I imagined, "What if there was a tool that turns text into something more engaging such as a video, wouldn't it improve my learning experience?" I created a prototype web application that takes text as an input and generates a video as an output. I plan to further work on the project targeting young college students who are aged between 18 to 23 because they tend to prefer learning through videos over books based on the survey I found. The technologies I used for the project are HTML, CSS, Javascript, Node.js, CCapture.js, ffmpegserver.js, Amazon Polly, Python, Flask, gevent, spaCy, and Pixabay API.

A multimodal AI story teller, built with Stable Diffusion, GPT, and neural text-to-speech (TTS). Given a prompt as an opening line of a story, GPT writes the rest of the plot; Stable Diffusion draws an image for each sentence; a TTS model narrates each line, resulting in a fully animated video of a short story, replete with audio and visuals. To develop locally, install dev dependencies and install pre-commit hooks. This will automatically trigger linting and code quality checks before each commit. The final video will be saved as /out/out.mp4, alongside other intermediate images, audio files, and subtitles. For more advanced use cases, you can also directly interface with Story Teller in Python code.

Implementation of Make-A-Video, new SOTA text to video generator from Meta AI, in Pytorch. They combine pseudo-3d convolutions (axial convolutions) and temporal attention and show much better temporal fusion. The pseudo-3d convolutions isn't a new concept. It has been explored before in other contexts, say for protein contact prediction as "dimensional hybrid residual networks". The gist of the paper comes down to, take a SOTA text-to-image model (here they use DALL-E2, but the same learning points would easily apply to Imagen), make a few minor modifications for attention across time and other ways to skimp on the compute cost, do frame interpolation correctly, get a great video model out. Passing in images (if one were to pretrain on images first), both temporal convolution and attention will be automatically skipped. In other words, you can use this straightforwardly in your 2d Unet and then port it over to a 3d Unet once that phase of the training is done.

Amiga Memories is a project (started & released in 2013) that aims to make video programmes that can be published on the internet. The images and sound produced by Amiga Memories are 100% automatically generated. The generator itself is implemented in Squirrel, the 3D rendering is done on GameStart 3D. An Amiga Memories video is mostly based on a narrative. The purpose of the script is to define the spoken and written content. The spoken text will be read by a voice synthesizer (Text To Speech or TTS), the written text is simply drawn on the image as subtitles. Here, in addition to the spoken & written narration, the script controls the camera movements as well as the LED activity of the computer. Amiga Memories' video images are computed by the GameStart 3D engine (pre-HARFANG 3D). Although the 3D assets are designed to be played back in real-time with a variable framerate, the engine is capable of breaking down the video sequence into the 30th or 60th of a second, as TGA files.

Implementation of NWT, audio-to-video generation, in Pytorch. The paper proposes a new discrete latent representation named Memcodes, which can be succinctly described as a type of multi-head hard-attention to learned memory (codebook) key/values. They claim the need for less codes and smaller codebook dimensions in order to achieve better reconstructions.

Implementation of Phenaki Video, which uses Mask GIT to produce text-guided videos of up to 2 minutes in length, in Pytorch. It will also combine another technique involving a token critic for potentially even better generations. A new paper suggests that instead of relying on the predicted probabilities of each token as a measure of confidence, one can train an extra critic to decide what to iteratively mask during sampling. This repository will also endeavor to allow the researcher to train on text-to-image and then text-to-video. Similarly, for unconditional training, the researcher should be able to first train on images and then fine tune on video.

Implementation of Video Diffusion Models, Jonathan Ho's new paper extending DDPMs to Video Generation - in Pytorch. Implementation of Video Diffusion Models, Jonathan Ho's new paper extending DDPMs to Video Generation - in Pytorch. It uses a special space-time factored U-net, extending generation from 2D images to 3D videos. 14k for difficult moving mnist (converging much faster and better than NUWA) - wip. Any new developments for text-to-video synthesis will be centralized at Imagen-pytorch. For conditioning on text, they derived text embeddings by first passing the tokenized text through BERT-large. You can also directly pass in the descriptions of the video as strings, if you plan on using BERT-base for text conditioning. This repository also contains a handy Trainer class for training on a folder of gifs. Each gif must be of the correct dimensions image_size and num_frames.

Video-subtitle-remover (VSR) is an AI-based software that removes hardcoded subtitles from videos or Pictures.

This is a collection of text-to-image tools, evolved from the artwork of the same name. Based on CLIP model and Lucent library, with FFT/DWT/RGB parameterizes (no-GAN generation). Illustrip (text-to-video with motion and depth) is added. DWT (wavelets) parameterization is added. Check also colabs below, with VQGAN and SIREN+FFM generators. Tested on Python 3.7 with PyTorch 1.7.1 or 1.8. Generating massive detailed textures, a la deepdream, fullHD/4K resolutions and above, various CLIP models (including multi-language from SBERT), continuous mode to process phrase lists (e.g. illustrating lyrics), pan/zoom motion with smooth interpolation. Direct RGB pixels optimization (very stable) depth-based 3D look (courtesy of deKxi, based on AdaBins), complex queries: text and/or image as main prompts, separate text prompts for style and to subtract (avoid) topics. Starting/resuming process from saved parameters or from an image.

This paper proposes a new GAN architecture for video generation with depth videos and color videos. The proposed model explicitly uses the information of depth in a video sequence as additional information for a GAN-based video generation scheme to make the model understands scene dynamics more accurately. The model uses pairs of color video and depth video for training and generates a video using the two steps. Generate the depth video to model the scene dynamics based on the geometrical information. To add appropriate color to the geometrical information of the scene, the domain translation from depth to color is performed for each image. This model has three networks in the generator. In addition, the model has two discriminators.

Implementation of NÜWA, state of the art attention network for text-to-video synthesis, in Pytorch. It also contains an extension into video and audio generation, using a dual decoder approach. It seems as though a diffusion-based method has taken the new throne for SOTA. However, I will continue on with NUWA, extending it to use multi-headed codes + hierarchical causal transformer. I think that direction is untapped for improving on this line of work. In the paper, they also present a way to condition the video generation based on segmentation mask(s). You can easily do this as well, given you train a VQGanVAE on the sketches beforehand. Then, you will use NUWASketch instead of NUWA, which can accept the sketch VAE as a reference. This repository will also offer a variant of NUWA that can produce both video and audio. For now, the audio will need to be encoded manually.

Implementation of Recurrent Interface Network (RIN), for highly efficient generation of images and video without cascading networks, in Pytorch. The author unawaredly reinvented the induced set-attention block from the set transformers paper. They also combine this with the self-conditioning technique from the Bit Diffusion paper, specifically for the latents. The last ingredient seems to be a new noise function based around the sigmoid, which the author claims is better than cosine scheduler for larger images. The big surprise is that the generations can reach this level of fidelity. Will need to verify this on my own machine. Additionally, we will try adding an extra linear attention on the main branch as well as self-conditioning in the pixel space. The insight of being able to self-condition on any hidden state of the network as well as the newly proposed sigmoid noise schedule are the two main findings.