Abstract: We present a general-purpose framework for image modelling and vision tasks based on probabilistic frame prediction. Our approach unifies a broad range of tasks, from image segmentation, to novel view synthesis and video interpolation. We pair this framework with an architecture we term Transframer, which uses U-Net and Transformer components to condition on annotated context frames, and outputs sequences of sparse, compressed image features. Transframer is the state-of-the-art on a variety of video generation benchmarks, is competitive with the strongest models on few-shot view synthesis, and can generate coherent 30 second videos from a single image without any explicit geometric information. A single generalist Transframer simultaneously produces promising results on 8 tasks, including semantic segmentation, image classification and optical flow prediction with no task-specific architectural components, demonstrating that multi-task computer vision can be tackled using probabilistic image models. Our approach can in principle be applied to a wide range of applications that require learning the conditional structure of annotated image-formatted data.
Transframer - Pytorch (wip)
Implementation of Transframer, Deepmind's U-net + Transformer architecture for up to 30 seconds video generation, in Pytorch
The gist of the paper is the usage of a Unet as a multi-frame encoder, along with a regular transformer decoder cross attending and predicting the rest of the frames. The author builds upon his prior work where images are encoded as sparse discrete cosine transform (DCT) sequences.
I will deviate from the implementation in this paper, using a hierarchical autoregressive transformer, and just a regular resnet block in place of the NF-net block (this design choice is just Deepmind reusing their own code, as NF-net was developed at Deepmind by Brock et al).
Update: On further meditation, there is nothing new in this paper except for generative modeling on DCT representations
Appreciation
- This work would not be possible without the generous sponsorship from Stability AI, as well as my other sponsors
Todo
- figure out if dct can be directly extracted from images in jpeg format
Citations
@article{Nash2022TransframerAF,
title = {Transframer: Arbitrary Frame Prediction with Generative Models},
author = {Charlie Nash and Jo{\~a}o Carreira and Jacob Walker and Iain Barr and Andrew Jaegle and Mateusz Malinowski and Peter W. Battaglia},
journal = {ArXiv},
year = {2022},
volume = {abs/2203.09494}
}
