We are a team focused on automating the training process of stable diffusion models (LoRA, Lycoris, Pivotal Tuning, etc.).

Our goal is to automate the entire pipeline, including data acquisition, data filtering, training, step selection, and platform deployment, saving manpower and optimizing model training quality to the maximum.

Our team comprises a Ph.D. in Software Engineering, a Ph.D. candidate in Computer Vision, professionals in art and design, and several AI waifu enthusiasts.

We are a purely non-profit team, and all our work is completely open, without any form of charge.

If you are interested in our work, you can join https://discord.gg/TwdHJ42N72, we are glad to see some good ideas, technologies or suggestions.

We have conducted multiple iterations on Pivotal Tuning training technology (training with a LoRA and one or more pt files). The progress is documented in the following technical blogs:

• Overview of v1.4 Training Automation
• Result Analysis of v1.4 Training Automation
• v1.5 Automation is coming... :)

Models we have trained or hosted can be found at:

• CyberHarem - Huggingface, including datasets used for training
• narugo1992 - Civitai, where only a few models are selected for publication to avoid offense to those who insist on manually training models, lol

Quantitative analysis from the blog "Result Analysis of v1.4 Training Automation" shows that v1.4 training automation has achieved a quite impressive level in both quality and quantity, but there's still room for further improvement. This is an ongoing effort for v1.5 and future versions.

Project Link: https://github.com/deepghs/imgutils

Project Documentation: https://deepghs.github.io/imgutils/main/index.html

This is a library for various common operations on anime images, including but not limited to:

• Tachie (Difference) Detection and Clustering
• Contrastive Character Image Pretraining
• Object Detection
• Edge Detection / Lineart Generation
• Monochrome Image Detection
• Image Rating
• Truncated Image Check
• Image Tagging
• Character Extraction

Check out the documentation for more features.

Project Link: https://github.com/deepghs/waifuc

A data pipeline framework based on dghs-imgutils, supporting:

• Fast data retrieval (local disk, danbooru, pixiv, zerochan, etc.)
• Swift data filtering (comic exclusion, monochrome image exclusion, multi-character image exclusion, irrelevant character exclusion, etc.)
• Rapid data saving (local, cloud; with metadata, saved in stable diffusion dataset format, etc.)
• Quick building of processing pipelines (connecting multiple aforementioned stages)

Note: This tool is currently a work in progress, although it's in use. It hasn't been released on PyPI and lacks comprehensive documentation. These aspects will be addressed soon.

We manage our models and datasets on Huggingface: https://huggingface.co/deepghs

In fact, our plans go beyond what's mentioned here. Other tools are continuously improving and will soon be released. Stay tuned!