# Multi-task

VisualCloze is a general-purpose image generation framework that learns through visual context, aiming to address the inefficiency of traditional task-specific models under diverse needs. This framework not only supports multiple internal tasks but can also generalize to unseen tasks, using visual examples to help the model understand the task. This approach leverages the strong generative priors of advanced image inpainting models, providing strong support for image generation.

Wan2GP is an improved version based on Wan2.1, aiming to provide an efficient and low-memory video generation solution for low-configuration GPU users. The model, through optimized memory management and accelerated algorithms, enables ordinary users to quickly generate high-quality video content on consumer-grade GPUs. It supports multiple tasks, including text-to-video, image-to-video, and video editing, and features a powerful video VAE architecture capable of efficiently handling 1080P videos. The emergence of Wan2GP lowers the barrier to entry for video generation technology, allowing more users to easily learn and apply it to real-world scenarios.

Wan2.1 is an open-source, advanced, large-scale video generation model designed to push the boundaries of video generation technology. Through innovative spatio-temporal variational autoencoders (VAEs), scalable training strategies, large-scale data construction, and automated evaluation metrics, it significantly improves model performance and versatility. Wan2.1 supports multiple tasks, including text-to-video, image-to-video, and video editing, and can generate high-quality video content. The model has demonstrated superior performance in several benchmark tests, even surpassing some closed-source models. Its open-source nature allows researchers and developers to freely use and extend the model, making it suitable for various applications.

Llama-3.1-Tulu-3-8B-SFT is part of the Tülu3 model family, a leading class of instruction-following models that offers fully open-source data, code, and guidelines aimed at modern post-training techniques. The model excels not only in chat tasks but also demonstrates outstanding performance across various tasks such as MATH, GSM8K, and IFEval.

LEO is a multimodal, multi-task all-in-one agent based on a large language model, capable of perceiving, localizing, reasoning, planning, and executing tasks in the 3D world. LEO achieves this through two stages of training: (i) 3D visual-language alignment and (ii) 3D visual-language action instruction tuning. We carefully curated and generated a large-scale dataset with object-level and scene-level multimodal tasks, requiring deep understanding and interaction with the 3D world. Through rigorous experiments, we demonstrate LEO's outstanding performance across a wide range of tasks, including 3D captioning, question answering, reasoning, navigation, and robot manipulation."