---
base_model: Qwen/Qwen3-VL-4B-Instruct
library_name: transformers
model_name: view2space_4b
tags:
- view2space
- eccv-2026
- multi-view-reasoning
- spatial-reasoning
- visual-reasoning
licence: license
---

# VIEW2SPACE: Studying Multi-View Visual Reasoning from Sparse Observations

`view2space_4b` is an ECCV 2026 VIEW2SPACE model built on top of
[Qwen/Qwen3-VL-4B-Instruct](https://huggingface.co/Qwen/Qwen3-VL-4B-Instruct).
It is designed for grounded multi-view visual reasoning from sparse
observations.

![VIEW2SPACE teaser](./teaser.png)

## Quick start

Please see the VIEW2SPACE GitHub repository for evaluation code and usage:

🧑‍💻[GitHub Repository](https://github.com/pokerme7777/VIEW2SPACE)

👀[Project Page](https://pokerme7777.github.io/VIEW2SPACE/)

## Quick links

- [Paper](http://arxiv.org/abs/2603.16506)
- [GitHub Repository](https://github.com/pokerme7777/VIEW2SPACE)
- [Public Testing Set](https://huggingface.co/datasets/Pokerme/view2space-v1)
- [Evaluation Code](https://github.com/pokerme7777/VIEW2SPACE/tree/main/src/eval)

## Overview

VIEW2SPACE studies how vision-language models reason across sparse and
heterogeneous viewpoints. Instead of solving a task from a single image, the
model must integrate partial observations from multiple views to form a more
complete spatial understanding.

This checkpoint is the `Qwen3-VL-4B` VIEW2SPACE model release and is intended
for multi-view visual reasoning under sparse observations.

## Model Summary

- Model name: `view2space_4b`
- Base model: `Qwen/Qwen3-VL-4B-Instruct`
- Architecture: `Qwen3VLForConditionalGeneration`
- Project: VIEW2SPACE
- Use case: multi-view visual reasoning from sparse observations
- Venue: ECCV 2026

## Highlights

- Built for grounded multi-view reasoning rather than single-image prediction.
- Targets sparse observations and heterogeneous viewpoints.
- Released together with the public VIEW2SPACE testing set and evaluation code.

## Resources

- Public testing release: [`view2space-v1`](https://huggingface.co/datasets/Pokerme/view2space-v1)
- Official repository: `https://github.com/pokerme7777/VIEW2SPACE`
- Public eval pipeline: `src/eval` in the official repository

## Usage Notes

- Use the official VIEW2SPACE repository for evaluation scripts and prompt formatting.
- The current public testing release is `view2space-v1`.
- If you need another public data format, please open an issue in the GitHub repository.


### Framework versions

- TRL: 0.26.2
- Transformers: 4.57.0
- Pytorch: 2.7.1+cu126
- Datasets: 4.4.2
- Tokenizers: 0.22.1

## Citations

```bibtex
@article{ke2026view2space,
  title={VIEW2SPACE: Studying Multi-View Visual Reasoning from Sparse Observations},
  author={Ke, Fucai and Cai, Zhixi and Li, Boying and Chen, Long and Lin, Beibei and Wang, Weiqing and Haghighi, Pari Delir and Haffari, Gholamreza and Rezatofighi, Hamid},
  journal={arXiv preprint arXiv:2603.16506},
  year={2026}
}
```