---
title: GIM-World
canonical_url: "https://www.modelscope.cn/models/nagara214/GIM-World"
md_url: "https://www.modelscope.cn/models/nagara214/GIM-World.md"
repository: nagara214/GIM-World
last_updated: 2026-09-11
license: apache-2.0
pipeline_tag: image-to-video
tasks:
  - image-to-video
base_model:
  - Wan-AI/Wan2.1-T2V-1.3B
base_model_relation: finetune
parameters: 3.1B
tensor_type:
  - F32
library_name:
  - safetensors
  - diffusers
downloads: 219
stars: 0
tags:
  - world-model
  - video-generation
  - memory
---

# GIM-World

> GIM-World - nagara214 在 ModelScope 开源的模型。GIM-World: Geometry-Aware Implicit Memory for Video World Models

nagara214/GIM-World 是 ModelScope 魔搭社区上的 3.1B 参数image-to-video模型，采用 apache-2.0 许可，基于 Wan-AI/Wan2.1-T2V-1.3B 构建。

- **Repository**: nagara214/GIM-World
- **License**: apache-2.0
- **Tasks**: image-to-video
- **Parameters**: 3.1B
- **Base model**: Wan-AI/Wan2.1-T2V-1.3B
- **Tags**: world-model, video-generation, memory
- **Downloads**: 219
- **Stars**: 0
- **Last updated**: 2026-09-11

Source: https://www.modelscope.cn/models/nagara214/GIM-World

---

# GIM-World: Geometry-Aware Implicit Memory for Video World Models

Checkpoints for **GIM-World** (SIGGRAPH Asia 2026).
Code: https://github.com/nagara214/GIM-World · Paper: https://arxiv.org/abs/2606.02436 · Project page: https://gim-world.github.io/

| Folder | Trained on | Notes |
|---|---|---|
| `first_person/` | MIND first-person split | camera = actor pose |
| `third_person/` | MIND third-person split | camera = chase camera pose |

Each folder contains

```
config.json                    # inference hyper-parameters read by GIMWorldPipeline
transformer/                   # Wan2.1-1.3B DiT fine-tuned with memory conditioning
memory_encoder.safetensors     # implicit memory encoder (2 blocks, compact stride 2)
camera_proj.safetensors        # per-block camera injection
action_embedding.safetensors   # action embedding
```

The Wan2.1 VAE and umT5 text encoder are loaded from `Wan-AI/Wan2.1-T2V-1.3B`.

```python
from gim import GIMWorldPipeline
pipe = GIMWorldPipeline.from_pretrained("GIM-World/first_person", "Wan2.1-T2V-1.3B")
result = pipe("path/to/clip_dir")   # video.mp4 + action.json
```

```bibtex
@article{wei2026gim,
  title   = {Geometry-Aware Implicit Memory for Video World Models},
  author  = {Wei, Zhengxuan and Guo, Xu and Li, Xinghui and Xiang, Xunzhi and Wei, Min and Zhu, Yiran and Wang, Qiulin and Wang, Xintao and Wan, Pengfei and Hou, Xiangwang and Fan, Qi},
  journal = {arXiv preprint arXiv:2606.02436},
  year    = {2026}
}
```