#NeurIPS 2023

Explore a meticulously curated repository focused on Vision Transformer and Attention, featuring comprehensive resources like papers, codes, and links to relevant websites. Maintained by Min-Hung Chen, this updated list invites contributions to enhance its comprehensiveness and includes the latest developments from major conferences such as NeurIPS 2023 and ICCV 2023. Researchers can contribute by opening issues or creating pull requests for any missed papers, ensuring a continually relevant resource for academic and enthusiast communities alike.

3D-OVS offers a PyTorch-based method for 3D scene segmentation using open vocabulary texts, eliminating the need for segmentation annotations. Compatible with Ubuntu 20.04, it leverages CLIP features for accurate segmentation using about 14GB of GPU memory. Comprehensive guidance is available for troubleshooting, text prompt refinement, and custom dataset handling, enhancing boundary detection and stability.

VisorGPT utilizes generative pre-training to enhance visual data comprehension. Presented at NeurIPS 2023, it offers tools like ControlNet and GLIGEN to improve image generation capabilities. Explore its features via Hugging Face and GitHub demos, with straightforward setup instructions. Open-source code and data emphasize contributions towards collaborative AI development in visual processing.

Explore UniControl, a unified diffusion model enabling controllable visual generation for various tasks within one framework. It achieves pixel-level precision by combining visual conditions for structural guidance and language prompts for style. By leveraging pretrained text-to-image models and a task-specific HyperNet, UniControl efficiently handles diverse condition-to-image tasks. This framework outperforms single-task models of similar sizes, representing a significant advancement in visual generation. Access includes open-source code, model checkpoints, and datasets for further exploration.

One-2-3-45 presents a novel approach in utilizing 2D diffusion models for 3D AI content generation with a forward-only paradigm that minimizes optimization time. It allows efficient creation of 3D models, as showcased by updates such as rendering scripts and APIs for inference. Accepted at NeurIPS 2023 and integrated with Hugging Face Spaces, the project offers revolutionary methods for 3D modeling with easy setup options. Interactive demos and model training using the Objaverse-LVIS dataset enhance user engagement.

OpenShape provides advanced 3D shape representation with zero-shot classification. Featuring publicly available training resources, it supports tasks like point cloud captioning and image generation using models such as PointBERT and SparseConv. Explore its contributions in 3D understanding through live demos and research.

This initiative presents a structured framework for developing, fine-tuning, and evaluating Large Language Models (LLMs) aimed at the financial sector. It grants open access to financial LLMs, instructional tuning datasets, and comprehensive datasets across diverse tasks, promoting transparency and collaborative research efforts. The project emphasizes multi-task capabilities and accommodates multi-modal financial data, enhancing understanding and prediction within the field of financial NLP.

Seamless is an AI-driven multilingual and multimodal translation project that supports extensive language coverage and promotes authentic interactions. The SeamlessM4T model underpins the broader ecosystem, including SeamlessExpressive and SeamlessStreaming, allowing for real-time, expressive translations and simultaneous translation capabilities in around 100 languages. Seamless utilizes the innovative UnitY2 architecture to improve translation efficiency and reduce latency.