# Autoregressive Model

Aurora is a next-generation autoregressive image generation model launched by Grok, trained on billions of internet samples, enabling a profound understanding of the world. Aurora excels in photo-realistic rendering and accurately follows text instructions, supporting multimodal input to draw inspiration from user-provided images or directly edit them. The new features are already available in selected countries on the ?? platform and will be rolled out to all users within a week.

OLMo 2 13B is a transformer-based autoregressive language model developed by the Allen Institute for AI (AI2), focusing on English academic benchmark testing. During training, it utilized up to 50 trillion tokens, demonstrating performance comparable to or even superior to similarly sized open models, and competing with the open-weight models from Meta and Mistral on English academic benchmarks. The release of OLMo 2 13B includes all code, checkpoints, logs, and relevant training details, aimed at advancing scientific research in language models.

Open-MAGVIT2 is an open-source series of autoregressive image generation models released by Tencent ARC Lab, featuring models ranging from 300M to 1.5B parameters. This project reproduces Google's MAGVIT-v2 tokenizer and achieves state-of-the-art reconstruction performance with a rFID of 1.17 on the ImageNet 256×256 dataset. By introducing asymmetric tokenization techniques, it decomposes large vocabularies into sub-vocabularies of varying sizes and enhances inter-token interaction through 'next sub-token prediction' to improve generation quality. All models and code are open-source, aimed at advancing innovation and creativity in the field of autoregressive visual generation.

Lumina-mGPT is part of a family of multimodal autoregressive models capable of performing various visual and language tasks, particularly excelling at generating flexible, realistic images from text descriptions. This model is implemented based on the xllmx module and supports LLM-centric multimodal tasks, making it suitable for in-depth exploration and rapid familiarization with the model's capabilities.

LlamaGen is a new family of image generation models that applies the original next token prediction paradigm of large language models to the field of visual generation. This model achieves state-of-the-art image generation performance without inductive bias on visual signals through appropriate extensions. LlamaGen re-examines the design space of image partitioners, the scalability properties of image generation models, and the quality of their training data.

VAR is a novel visual autoregressive modeling method that surpasses diffusion models, achieving more efficient image generation. It establishes the power-law scaling laws of visual generation and possesses zero-shot generalization capabilities. VAR provides a range of pre-trained models of different sizes for users to explore and utilize.