README.md

---
license: cc-by-nc-4.0
language:
  - en
library_name: transformers
pipeline_tag: image-text-to-text
tags:
  - vision-language-model
  - multimodal
  - panoramic-understanding
  - 360-degree
  - equirectangular-panorama
  - spatial-reasoning
  - panoworld
base_model:
  - Qwen/Qwen3.5-9B
datasets:
  - wcccp/Pano_dataset
---


# PanoWorld-Hstar

PanoWorld-Hstar is a vision-language model based on **Qwen3.5-9B**, developed for 360-degree panoramic understanding and spatial reasoning.

The model is part of the **PanoWorld** project, which focuses on ERP-native panoramic perception, global spatial topology understanding, and human-centric visual search in 360° scenes.

* Project: https://github.com/wcpcp/PanoWorld
* Model: https://huggingface.co/wcccp/PanoWorld
* Dataset: https://huggingface.co/datasets/wcccp/Pano_dataset

## Model Description

PanoWorld-Hstar is fine-tuned for vision-language understanding in equirectangular panorama images. It is designed to improve model capability on panoramic scene captioning, spatial relation reasoning, direction understanding, and 360° visual question answering.

## Intended Use

This model is intended for research on:

* 360° panoramic image understanding
* panoramic visual question answering
* spatial and directional reasoning
* human-centric visual search in panoramic scenes
* embodied AI and panoramic scene perception

## Usage

```python
import torch
from transformers import AutoProcessor, Qwen3_5ForConditionalGeneration

model_id = "wcccp/PanoWorld-Hstar"

processor = AutoProcessor.from_pretrained(model_id, trust_remote_code=True)
model = Qwen3_5ForConditionalGeneration.from_pretrained(
    model_id,
    torch_dtype=torch.bfloat16,
    device_map="auto",
    trust_remote_code=True,
)

messages = [
    {
        "role": "user",
        "content": [
            {"type": "image", "image": "example_panorama.jpg"},
            {"type": "text", "text": "Describe this 360-degree panoramic scene."},
        ],
    }
]

inputs = processor.apply_chat_template(
    messages,
    tokenize=True,
    add_generation_prompt=True,
    return_dict=True,
    return_tensors="pt",
).to(model.device)

generated_ids = model.generate(
    **inputs,
    max_new_tokens=512,
)

generated_ids_trimmed = [
    output_ids[len(input_ids):]
    for input_ids, output_ids in zip(inputs.input_ids, generated_ids)
]

response = processor.batch_decode(
    generated_ids_trimmed,
    skip_special_tokens=True,
    clean_up_tokenization_spaces=False,
)[0]

print(response)
```

Please use a recent version of `transformers` that supports Qwen3.5.


## Citation

```bibtex
@misc{wang2026panoworld,
      title={PanoWorld: Towards Spatial Supersensing in 360$^\circ$ Panorama World}, 
      author={Changpeng Wang and Xin Lin and Junhan Liu and Yuheng Liu and Zhen Wang and Donglian Qi and Yunfeng Yan and Xi Chen},
      year={2026},
      eprint={2605.13169},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2605.13169}, 
}
```