Tier-Flow/LoopVL

Recurrent visual computation for vision-language models Paper  ·  🤗 Hugging Face  ·  ModelScope  ·  Getting started  ·  Benchmarks  ·  Experiments LoopVL is a vision-language model that reuses Transformer modules across recurrent computation. LoopVL-1B follows the L → L → L → H → L → L → L → H schedule, with 16 layers per module call and 128 effective layer applications per forward pass. This repository brings together image-question answering, benchmark evaluation, and experiments for studying visual attention and hidden-state dynamics. Model weights are available on Hugging Face and ModelScope as a single model.safetensors file, accompanied by configuration and tokenizer files. Use Linux, Python 3.12, and an NVIDIA GPU for inference and evaluation. Install a CUDA-compatible PyTorch/torchvision pair for your system; the reference environment uses PyTorch 2.12.1 and torchvision 0.27.1. Then install the project dependencies: Run the commands below from the repository root. Choose either model hub and download…