DeepMind joins forces with MIT to develop Fluid to allow autoregressive models to achieve large-scale scaling of literate graphs

DeepMind and MIT researchers jointly developed the Fluid model to enable the extension of autoregressive models for text-to-image generation. The model uses continuous Token and random order strategies to improve visual quality and evaluation metrics and reduce information loss.Fluid scores well in several benchmarks, but there is still room for improvement in positional relationship understanding and rendering two-object tasks. This research may help close the gap between vision and language models in terms of scaling performance. (DeepTech)