🎬 One-Minute Video Generation via Test-Time Transformer Training

Researchers introduced Test-Time Training (TTT) layers to enhance the ability of pre-trained Diffusion Transformers to generate longer, more complex videos from text. These novel layers, inspired by meta-learning, allow the model's hidden states to adapt during the video generation process. To validate their approach, they created a dataset of annotated Tom and Jerry cartoons for training and evaluation. Their model, incorporating TTT layers, outperformed existing methods in generating coherent, minute-long videos with multi-scene stories and dynamic motion, as judged by human evaluators. While promising, the generated videos still exhibit some artifacts, and the method's efficiency could be improved. The study demonstrates a step forward in creating longer, story-driven videos from textual descriptions.