Open Source 🎲 RandAR: Decoder-only Autoregressive Visual Generation in Random Orders

[Project Page] [arXiv] [HuggingFace]

Overview

Ever thinking about what is the prerequisite for a visual model achieving the impact of GPT in language? The prequisite should be its ability of zero-shot generalization to various applications, prompts, etc. Our RandAR is one of the attempts towards this objective.

🎲 RandAR is a decoder-only AR model generating image tokens in arbitrary orders.

🚀 RandAR supports parallel-decoding without additional fine-tuning and brings 2.5