README.md

---
license: gpl-3.0
base_model:
- Wan-AI/Wan2.1-T2V-14B
- Wan-AI/Wan2.1-T2V-1.3B
pipeline_tag: text-to-video
---

# Flow Oil - an oil painting lora for wan


## 模型介绍

这是一个由时辰和海辛以及阿文一起，以经典油画为基础，并参考了梵高的流动特点进行训练的艺术化WAN14B LoRA的预览版。模型能够将现实世界的视觉元素转换为油画风格的表现形式，例如将爆炸、灯光等效果转化为离散的油画笔触表示。

更全面的版本以及支持图生视频的版本还在训练中。该模型由海辛和阿文提供想法和支持，在这个模型里我们发现WAN可以很好地学习到训练素材对于世界特征的描述，呈现出令人惊喜的艺术效果。

## 模型列表

### 1. 14b_oil_motion_preview1.0.safetensors
- 使用33帧视频训练
- 特点：动态更足，颜色更丰富
- 适用场景：需要流畅动态效果的视频生成

**预览效果：**

<table style="width: 100%; text-align: left; margin-top: 20px;" border="0">
  <tbody>
    <tr>
      <td>
        <video loop autoplay controls width="100%" src="https://huggingface.co/latentcat/flow-oil-WAN14BT2V/resolve/main/video/model1_1.mp4"></video>
      </td>
      <td>
        <video loop autoplay controls width="100%" src="https://huggingface.co/latentcat/flow-oil-WAN14BT2V/resolve/main/video/model1_2.mp4"></video>
      </td>
      <td>
        <video loop autoplay controls width="100%" src="https://huggingface.co/latentcat/flow-oil-WAN14BT2V/resolve/main/video/model1_3.mp4"></video>
      </td>
    </tr>
    <tr>
      <td>
        <video loop autoplay controls width="100%" src="https://huggingface.co/latentcat/flow-oil-WAN14BT2V/resolve/main/video/model1_4.mp4"></video>
      </td>
      <td>
        <video loop autoplay controls width="100%" src="https://huggingface.co/latentcat/flow-oil-WAN14BT2V/resolve/main/video/model1_5.mp4"></video>
      </td>
      <td></td>
    </tr>
  </tbody>
</table>

### 2. 14b_oil_style_preview1.0.safetensors
- 分两次训练再合并
- 特点：在480p分辨率上对风格的保持最好，720P效果较弱
- 适用场景：中低分辨率的艺术风格转换

**预览效果：**

<table style="width: 100%; text-align: left; margin-top: 20px;" border="0">
  <tbody>
    <tr>
      <td>
        <video loop autoplay controls width="100%" src="https://huggingface.co/latentcat/flow-oil-WAN14BT2V/resolve/main/video/model2_1.mp4"></video>
      </td>
      <td>
        <video loop autoplay controls width="100%" src="https://huggingface.co/latentcat/flow-oil-WAN14BT2V/resolve/main/video/model2_2.mp4"></video>
      </td>
      <td>
        <video loop autoplay controls width="100%" src="https://huggingface.co/latentcat/flow-oil-WAN14BT2V/resolve/main/video/model2_3.mp4"></video>
      </td>
    </tr>
    <tr>
      <td>
        <video loop autoplay controls width="100%" src="https://huggingface.co/latentcat/flow-oil-WAN14BT2V/resolve/main/video/model2_4.mp4"></video>
      </td>
      <td>
        <video loop autoplay controls width="100%" src="https://huggingface.co/latentcat/flow-oil-WAN14BT2V/resolve/main/video/model2_5.mp4"></video>
      </td>
      <td></td>
    </tr>
  </tbody>
</table>

### 3. 14b_oil_style_v2_preview1.0.safetensors
- 带有720P分辨率图片并进行了更多步数的训练
- 特点：高分辨率风格保持更好
- 适用场景：需要高清晰度的油画风格视频

**预览效果：**

<table style="width: 100%; text-align: left; margin-top: 20px;" border="0">
  <tbody>
    <tr>
      <td>
        <video loop autoplay controls width="100%" src="https://huggingface.co/latentcat/flow-oil-WAN14BT2V/resolve/main/video/model3_1.mp4"></video>
      </td>
      <td>
        <video loop autoplay controls width="100%" src="https://huggingface.co/latentcat/flow-oil-WAN14BT2V/resolve/main/video/model3_2.mp4"></video>
      </td>
      <td>
        <video loop autoplay controls width="100%" src="https://huggingface.co/latentcat/flow-oil-WAN14BT2V/resolve/main/video/model3_3.mp4"></video>
      </td>
    </tr>
    <tr>
      <td>
        <video loop autoplay controls width="100%" src="https://huggingface.co/latentcat/flow-oil-WAN14BT2V/resolve/main/video/model3_4.mp4"></video>
      </td>
      <td>
        <video loop autoplay controls width="100%" src="https://huggingface.co/latentcat/flow-oil-WAN14BT2V/resolve/main/video/model3_5.mp4"></video>
      </td>
      <td></td>
    </tr>
  </tbody>
</table>

## 提示词格式

```
Oil painting style, VanGogh, VanGogh style. {prompt}. Impasto oil painting in the style of Van Gogh's, impressionistic painting, oil painting, loose brush strokes, canvas texture, impasto technique, Van Gogh style.
```

## 惊喜
我们发现在其他场景上出现了巨大的泛化能力，请看炸月球vcr
<table style="width: 100%; text-align: left; margin-top: 20px;" border="0">
  <tbody>
    <tr>
      <td>
        <video loop autoplay controls width="100%" src="https://huggingface.co/latentcat/flow-oil-WAN14BT2V/resolve/main/video/moon_1.mp4"></video>
      </td>
      <td>
        <video loop autoplay controls width="100%" src="https://huggingface.co/latentcat/flow-oil-WAN14BT2V/resolve/main/video/moon_2.mp4"></video>
      </td>
      <td>
        <video loop autoplay controls width="100%" src="https://huggingface.co/latentcat/flow-oil-WAN14BT2V/resolve/main/video/moon_3.mp4"></video>
      </td>
    </tr>
    <tr>
      <td>
        <video loop autoplay controls width="100%" src="https://huggingface.co/latentcat/flow-oil-WAN14BT2V/resolve/main/video/moon_4.mp4"></video>
      </td>
      <td></td>
      <td></td>
    </tr>
  </tbody>
</table>
## 使用方法

1. 下载对应的模型文件
2. 将模型加载到支持WAN14B的视频生成框架中
3. 根据需要的效果选择相应的模型：
   - 需要动态效果优先选择 `14b_oil_motion_preview1.0`
   - 480p分辨率下使用 `14b_oil_style_preview1.0`
   - 720p高分辨率场景使用 `14b_oil_style_v2_preview1.0`

## 已知问题与限制

- 预览版可能在某些复杂场景下表现不稳定
- 高分辨率生成时可能需要更多的计算资源
- 某些特定颜色和纹理可能需要额外的提示词调整
- 使用 SLG 的时候边缘会出现波动

## 未来计划

- 发布完整版模型，支持更多场景和风格变化
- 提供专门针对图生视频优化的版本
- 增加更多艺术风格的支持

## 致谢

感谢海辛和阿文提供的想法和支持，使这个模型成为可能。

---

# Flow Oil - an oil painting lora for wan

## Model Introduction

This is a preview version of an artistic WAN14B LoRA trained by Shichen, Haixin, and Awen, based on classic oil paintings and incorporating Van Gogh's flowing characteristics. The model can transform visual elements from the real world into oil painting style representations, such as converting explosions, lighting effects, etc. into discrete oil painting brush strokes.

A more comprehensive version and a version supporting image-to-video are still in training. This model was developed with ideas and support from Haixin and Awen. In this model, we discovered that WAN can effectively learn the description of world features from training materials, producing surprisingly artistic effects.

## Model List

### 1. 14b_oil_motion_preview1.0.safetensors
- Trained using 33 frames of video
- Features: More dynamic, richer colors
- Suitable scenarios: Video generation requiring smooth dynamic effects

**Preview effects:**

<table style="width: 100%; text-align: left; margin-top: 20px;" border="0">
  <tbody>
    <tr>
      <td>
        <video loop autoplay controls width="100%" src="https://huggingface.co/latentcat/flow-oil-WAN14BT2V/resolve/main/video/model1_1.mp4"></video>
      </td>
      <td>
        <video loop autoplay controls width="100%" src="https://huggingface.co/latentcat/flow-oil-WAN14BT2V/resolve/main/video/model1_2.mp4"></video>
      </td>
      <td>
        <video loop autoplay controls width="100%" src="https://huggingface.co/latentcat/flow-oil-WAN14BT2V/resolve/main/video/model1_3.mp4"></video>
      </td>
    </tr>
    <tr>
      <td>
        <video loop autoplay controls width="100%" src="https://huggingface.co/latentcat/flow-oil-WAN14BT2V/resolve/main/video/model1_4.mp4"></video>
      </td>
      <td>
        <video loop autoplay controls width="100%" src="https://huggingface.co/latentcat/flow-oil-WAN14BT2V/resolve/main/video/model1_5.mp4"></video>
      </td>
      <td></td>
    </tr>
  </tbody>
</table>

### 2. 14b_oil_style_preview1.0.safetensors
- Trained twice and then merged
- Features: Best style preservation at 480p resolution, weaker effect at 720P
- Suitable scenarios: Artistic style transformation at medium to low resolutions

**Preview effects:**

<table style="width: 100%; text-align: left; margin-top: 20px;" border="0">
  <tbody>
    <tr>
      <td>
        <video loop autoplay controls width="100%" src="https://huggingface.co/latentcat/flow-oil-WAN14BT2V/resolve/main/video/model2_1.mp4"></video>
      </td>
      <td>
        <video loop autoplay controls width="100%" src="https://huggingface.co/latentcat/flow-oil-WAN14BT2V/resolve/main/video/model2_2.mp4"></video>
      </td>
      <td>
        <video loop autoplay controls width="100%" src="https://huggingface.co/latentcat/flow-oil-WAN14BT2V/resolve/main/video/model2_3.mp4"></video>
      </td>
    </tr>
    <tr>
      <td>
        <video loop autoplay controls width="100%" src="https://huggingface.co/latentcat/flow-oil-WAN14BT2V/resolve/main/video/model2_4.mp4"></video>
      </td>
      <td>
        <video loop autoplay controls width="100%" src="https://huggingface.co/latentcat/flow-oil-WAN14BT2V/resolve/main/video/model2_5.mp4"></video>
      </td>
      <td></td>
    </tr>
  </tbody>
</table>

### 3. 14b_oil_style_v2_preview1.0.safetensors
- Trained with 720P resolution images and more training steps
- Features: Better style preservation at high resolution
- Suitable scenarios: Oil painting style videos requiring high definition

**Preview effects:**

<table style="width: 100%; text-align: left; margin-top: 20px;" border="0">
  <tbody>
    <tr>
      <td>
        <video loop autoplay controls width="100%" src="https://huggingface.co/latentcat/flow-oil-WAN14BT2V/resolve/main/video/model3_1.mp4"></video>
      </td>
      <td>
        <video loop autoplay controls width="100%" src="https://huggingface.co/latentcat/flow-oil-WAN14BT2V/resolve/main/video/model3_2.mp4"></video>
      </td>
      <td>
        <video loop autoplay controls width="100%" src="https://huggingface.co/latentcat/flow-oil-WAN14BT2V/resolve/main/video/model3_3.mp4"></video>
      </td>
    </tr>
    <tr>
      <td>
        <video loop autoplay controls width="100%" src="https://huggingface.co/latentcat/flow-oil-WAN14BT2V/resolve/main/video/model3_4.mp4"></video>
      </td>
      <td>
        <video loop autoplay controls width="100%" src="https://huggingface.co/latentcat/flow-oil-WAN14BT2V/resolve/main/video/model3_5.mp4"></video>
      </td>
      <td></td>
    </tr>
  </tbody>
</table>

## Prompt Format

```
Oil painting style, VanGogh, VanGogh style. {prompt}. Impasto oil painting in the style of Van Gogh's, impressionistic painting, oil painting, loose brush strokes, canvas texture, impasto technique, Van Gogh style.
```

## Surprise
We discovered tremendous generalization capabilities in other scenarios, check out the moon explosion vcr
<table style="width: 100%; text-align: left; margin-top: 20px;" border="0">
  <tbody>
    <tr>
      <td>
        <video loop autoplay controls width="100%" src="https://huggingface.co/latentcat/flow-oil-WAN14BT2V/resolve/main/video/moon_1.mp4"></video>
      </td>
      <td>
        <video loop autoplay controls width="100%" src="https://huggingface.co/latentcat/flow-oil-WAN14BT2V/resolve/main/video/moon_2.mp4"></video>
      </td>
      <td>
        <video loop autoplay controls width="100%" src="https://huggingface.co/latentcat/flow-oil-WAN14BT2V/resolve/main/video/moon_3.mp4"></video>
      </td>
    </tr>
    <tr>
      <td>
        <video loop autoplay controls width="100%" src="https://huggingface.co/latentcat/flow-oil-WAN14BT2V/resolve/main/video/moon_4.mp4"></video>
      </td>
      <td></td>
      <td></td>
    </tr>
  </tbody>
</table>

## Usage Instructions

1. Download the corresponding model file
2. Load the model into a video generation framework that supports WAN14B
3. Choose the appropriate model based on the desired effect:
   - For dynamic effects, prioritize `14b_oil_motion_preview1.0`
   - For 480p resolution, use `14b_oil_style_preview1.0`
   - For 720p high-resolution scenarios, use `14b_oil_style_v2_preview1.0`

## Known Issues and Limitations

- The preview version may perform unstably in some complex scenarios
- Higher resolution generation may require more computational resources
- Some specific colors and textures may require additional prompt adjustments
- When using SLG, edges may show fluctuations

## Future Plans

- Release the complete model version, supporting more scenarios and style variations
- Provide versions specifically optimized for image-to-video
- Add support for more artistic styles

## Acknowledgements

Thanks to Haixin and Awen for providing ideas and support, making this model possible.

---