之前我就一直在考虑通过开源的生图模型批量制造游戏资产，但是之前的那些模型控制力低，理解能力较差，无法较好地完成这个任务。而GPT image生成的游戏资产虽然一致性较好，但是它的噪点极多，实在让人难以忍受。

而Moss模型真的很让人惊叹，他使用了一个巨大的文本模型，这使得我们可以通过书写JSON来精确地控制构图。所以我训练了这个LoRA，这个LoRA是使用105张2.5D等距游戏场景图训练的，可以直接画游戏场景图和游戏资产，这算是一个实验性的东西，接下来我将尝试直接对MOSS模型本身进行一次微调，特化他的游戏资产生产能力，将其专门用于游戏资产生产。

这个LoRA的触发词为wakfu-iso

I had previously considered using open-source image generation models to mass-produce game assets; however, earlier models lacked sufficient control and comprehension, making them ill-suited for the task. Meanwhile, while game assets generated via GPT-based image models offered better consistency, they suffered from an intolerable level of noise.

The MOSS model, by contrast, is truly remarkable. It leverages a massive text-based model, allowing for precise control over composition through JSON-formatted input. Consequently, I trained a LoRA—using a dataset of 105 2.5D isometric game scene images—that is capable of generating both game environments and individual assets. This is an experimental project; my next step will be to fine-tune the MOSS model itself, specifically optimizing it for the creation of game assets.

The trigger word for this LoRA is wakfu-iso.