Kohya Trainer

Train LoRAs on Google Colab in one click. The notebooks run kohya-ss/sd-scripts, pinned to an exact upstream commit, so they keep working when upstream changes.

What it is for

kohya-trainer trains LoRAs on Colab the way LoRA trainers already work: sd-scripts names for every setting, the usual folder and caption habits, and the steps people take to get from images to a finished LoRA.

• The 1-click notebook is one press: add your images and train. Every setting has a default. It is not a cut-down trainer: its cell takes any sd-scripts setting, so a power user can use it too.
• The advanced notebook runs the same steps one cell at a time, so you can check and redo each stage.
• Nothing goes in unless it helps you finish a LoRA in one sitting.

Notebooks

Model	GPU	1-click	Advanced (per-stage)
Illustrious / NoobAI	L4
Anima	L4
MOSS-Image	L4

Other models

Model	GPU	1-click	Advanced (per-stage)
Stable Diffusion 1.5	L4
SDXL	L4

The notebooks ask Colab for an L4 GPU, which needs a paid Colab plan. Each notebook says how it runs on other GPUs.

1-click: pick a base model, point it at your images, and press run. You get the LoRA and sample images, and with save_every_n_epochs at 1 (the default) a file for each epoch. Any other sd-scripts setting goes in config_overrides in the cell's code.

Advanced: the same steps in separate cells (download, dataset, tagging, configure, train), so you can check and redo each stage.

Both can upload each saved epoch to the Hugging Face Hub and log to Weights & Biases.

Your dataset

• A folder, a .zip, a link to a .zip, or a Hugging Face dataset repo (user/name).
• Subfolders named like 10_mychar train with that many repeats each.
• Each image can have a .txt caption file with the same name. If no image has one, the WD tagger tags them. On MOSS-Image the tagger is off by default, so set trigger_word.

With Google Drive mounted (the default), the LoRA files go to MyDrive/lora/<project>/output.

Tokens

Add these to Colab Secrets when you need them: HF_TOKEN (read access for gated models, write access for uploads), WANDB_API_KEY and CIVITAI_TOKEN. The notebooks never write a token to disk or print it.

Contributing

See CONTRIBUTING.md.

Acknowledgements

• kohya-ss/sd-scripts does the training.
• The 2026 rebuild was written with Claude Code (Claude Opus 5.5).

License

Apache-2.0. The training backend, kohya-ss/sd-scripts, is also Apache-2.0.