ppbrown/onegirl200

---
license: openrail
tags:
- sdxl
---
# Current sample model

https://civitai.com/models/508420

The above is SDXL, and not very good. A better one is under way.

# Overview
This is my attempt at creating a truely open source SDXL model that people might be interested in using....
and perhaps copying the spirit and creating other open source models.
I'm including EVERYTHING I used to create my onegirl200 model:
* The images
* The captions
* The OneTrainer json preset file
* And my specific method i used to get here.

I've been playing around with the thousands of images I've filtered so far from danbooro, at
https://huggingface.co/datasets/ppbrown/danbooru-cleaned
So, the images here are a strict subset of those images.
I also used their tagging ALMOST as-is. I only added one tag: "anime"

See [METHODOLOGY-adamw.md] for a detailed description of what I personally did to coax a model out
of this dataset.

I also plan to try other training methods.

# Memory usage tips
I am using an RTX4090 card, which has 24 GB of VRAM. So I optimize for best quality, and then fastest speed,
that I can fit on my card.
Currently, that means bf16 SDXL or Cascade model finetunes, using "Default" attention, and no gradient saves.

You can save memory, at the sacrifice of speed, by enabling gradient saving. You can save more memory,
at the sacrifice of a little quality, by switching to Xformers attention.
Using those adjustments, you can run adafactor/adafactor finetunes on a 16GB card.

原始信息汇总

数据集概述

这是一个尝试创建一个真正开源的SDXL模型的项目，旨在吸引人们使用并可能以此为灵感创建其他开源模型。该项目包括创建onegirl200模型所使用的所有内容：

• 图像
• 描述
• OneTrainer json预设文件
• 具体的创建方法

图像数据是从ppbrown/danbooru-cleaned数据集中筛选出的一个严格子集。标签几乎直接使用，仅添加了一个“anime”标签。

详细的方法描述可以在[METHODOLOGY-adamw.md]文件中找到。

内存使用建议

该项目使用RTX4090显卡（24 GB VRAM）进行优化，以实现最佳质量和最快速度。当前配置为bf16 SDXL或Cascade模型微调，使用“Default”注意力，不保存梯度。

可以通过启用梯度保存来牺牲速度以节省内存，或通过切换到Xformers注意力来牺牲一点质量以节省更多内存。这些调整使得可以在16GB显卡上运行adafactor/adafactor微调。