通过 gpt image 2.5 修改产出，最终训练得到了这个 lora

This LoRA was trained using outputs modified via GPT Image 2.5.

使用了 30 张图片去训练，包含正面，侧面，背面，俯视，仰视

Thirty images were used for training, covering front, side, and back views, as well as top-down and low-angle views.

Both the chibi-style Grok-chan and the younger Grok-chan will share a single page; no separate pages will be created for them.

训练的时候忘记了现在是叫grok，而不是叫推特，所以打标和触发词都是“推特娘”

During training, I forgot that the name is grok now, not Twitter, so the label and trigger words are both "推特娘"

展示图大差不差都是因为懒，前两天DSH弄坏电脑后，原先存的pose和服装提示词都没了，这些都是ollama部署的模型批量出的。

The images shown are all pretty much the same—mostly out of laziness. A couple of days ago, DSH caused my computer to crash, and I lost all my saved prompts for poses and outfits. These were generated in batches using models deployed via Ollama.

推荐一下自己的批量加载提示词节点 https://github.com/update-loulan/comfyui-prompt-batch-writer

I’d like to share a custom node I created for batch-loading prompts: https://github.com/update-loulan/comfyui-prompt-batch-writer