Kohya Colabs
Accessible Google Colab notebooks for Stable Diffusion Lora training, based on the work of kohya-ss and Linaqruf.
If you like my work consider leaving me a tip :)
Lora making Guide | Click Here | - |
Stable Diffusion guide | Click Here | Click Aquí |
📊 Dataset Maker - Features
- Able to scrape hundreds of images from the popular anime gallery Gelbooru, that match the conditions set by the user.
- Finds duplicate images using the FiftyOne open-source software.
- Displays the user's dataset back to them through the FiftyOne interface so that they may manually curate their images.
- Able to generate tags for all your anime images using the Waifu Diffusion 1.4 Tagger model.
- Able to generate captions for all your images using the BLIP model.
- Gives you the ability to edit hundreds of text files at once, to add/remove/replace tags inside them dynamically.
- Works inside your Google Drive by default.
- Connects easily with Lora Trainer.
⭐ Lora Trainer - Features
- Can train LoRA, LoCon and LoHa.
- New feature: min-snr-gamma, optimizes loss to improve training efficiency.
- One click to install and start training.
- Offers all useful training parameters while keeping it simple and accessible.
- Helpful parameter descriptions and runtime messages.
- Can continue training an existing Lora (though not ideal).
- Advanced features in the form of custom configurations, allowing training with multiple datasets at once and more.
- Uses the latest technologies to load and train quickly.
- Utilizes kohya-ss scripts as a backend, an industry standard.
- Sources code from Linaqruf's colabs. Thank you!
- Works inside your Google Drive by default.
- Connects easily with Dataset Maker.