Preparing Dataset for Flux LoRA training

I'm a little bit sad about this answer, But anyway, Trying to find another solution...

Luma and Runway Test

This image(Figure 1) is taken from an ancient book and is a part of the training dataset for the latest Flux Woodblock prints style, Lora. I have decreased the resolution of this image from 768*1518 to 1280*720 in order to make it compatible with two of the most advanced and popular tools for AI artists, namely Luma and Runway Gen3.

The purpose of this process is to check whether it is possible to generate animation using those two online AI platforms by utilizing unusual data, such as unique illustration styles or non-realistic, non-3D software-generated images.

Figure1

Prompt: a mysterious, monkey-like creature with a long, curved tail and a slender body. As the animation begins, the creature suddenly leaps forward, dramatically extending its left arm as if trying to grasp something beyond reach. Its body twists sharply as if it's preparing to pounce or escape, while its expression shifts between cunning and curiosity. The tail swings widely in the air, following the dynamic motion of the creature

Runway Gen2 Test Result

Luma Animation Test Result

Other- Luma Blending Start and End Frame Test

Conclusion: the unusual dataset is not compatible with those online AI generative platforms so far, probably because the platform AI model only contained commonly open-sourced datasets that they can access for free and legally(although they can steal).

💡Reflection

Unique or customized styled artistic products still have commercial value, especially for The future with severe homogenization of artistic styles

---

Lora数据集（如图），教程链接

1.图片都是1024*1024 统一大小

2.图片的种类多样

3.现在的flux Lora可以打标也可以不打标

4.训练图片超过了50张

5.

Road map： 1.技术准备阶段： 训练lora - 动画- 三维

2.创作阶段：

剧本-storyboard-animatic-三维-合成-后期

3.宣发阶段： 投奖-各大平台宣传

先训练一个角色的flux lora

还有一个style 的lora

*然后单线条用toon craft 上色和搞中间帧试试