Tier-Flow / LoopVL

🚀 Try LoopVL for Free 
 Upload an image, ask a question, and explore different H/L recurrence schedules directly in your browser. 
 🤗 Launch the Interactive Demo → 
---
 Recurrent visual computation for vision-language models 
 Paper  · 
 🤗 Hugging Face  · 
 ModelScope  · 
 Getting started  · 
 Benchmarks  · 
 Experiments 
> [!TIP]
> For **LoopVL-1B**, we recommend **direct-answer prompting** for benchmark
> evaluation and exploring model behavior. **Chain-of-thought prompting is
> recommended only for mathematics tasks.**
>
> We will explore larger-parameter visual-loop models and alternative approaches
> to address the limitations observed in this generation.
## Reproduction prompt
 [code] 
## Overview
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.
## Getting started
### 1. Set up the environment
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:
 [code] 
Run the commands below from the repository root.
### 2. Download the model
Choose either model hub and download the complete snapshot into `model/`:
 [code] 
 [code] 
Check that the model files are ready:
 [code] 
The implementation lives in `runtime/`; weights, configuration, tokenizer and
image-processor settings live in `model/`. The LoopVL loader handles the
checkpoint's mixed BF16/FP32 precision automatically. See the
model setup guide for Python download alternatives and
the complete file layout.
### 3. Ask a question about an image
 [code] 
The answer is printed to the terminal. Generation uses greedy decoding, and
`--budget` sets the maximum number of generated tokens. Use `--model-dir` to
load a snapshot stored elsewhere.
For a mathematics question with step-by-step reasoning:
 [code] 
See the runtime guide for Python integration and model-loading
details.
## Benchmark evaluation
The evaluation suite covers **16 datasets** across general visual reasoning,
hallucination, mathematics and science. It also reports the POPE-A/P/R and
MMK12 subject-level scores.
### Prepare the data
 [code] 
The downloader retrieves the pinned public GitHub Release assets into
`data/server_snapshot_20260906/` and verifies both archive and per-file checksums.
### Run the suite
 [code] 
Choose one command for your hardware. `--gpus` selects the GPU indices and
`--workers-per-gpu` sets the maximum number of workers o