kohya-colab: LoRA Training Notebooks That Run in a Browser Tab

Two notebooks, not a training script

kohya-colab is a set of accessible Google Colab notebooks for Stable Diffusion LoRA training, and it is built on the work of kohya-ss and Linaqruf rather than being an original training implementation. That distinction matters. The upstream projects are command line toolkits that expect a prepared environment, a Python runtime and a GPU; the work these notebooks do is packaging that into something you can open from a link, press play on, and fill in from a form. The repository holds four notebooks in English: Dataset_Maker, Lora_Trainer, Lora_Trainer_XL and Lora_Trainer_XL_Legacy, plus two Spanish translations of the Dataset Maker and the Lora Trainer. Two helper Python files sit at the top level, `dracula.py` and the two `train_network_wrapper.py` variants, which is where the actual training entry points are wrapped for the notebook environment.

Dataset Maker handles the tedious part first

The first notebook is where most of the work happens, and its feature list is what makes the trainer usable. It can scrape hundreds of images from the anime gallery Gelbooru, matching conditions the user sets, which saves you from assembling a training set by hand. It finds duplicate images using the FiftyOne open-source software, then shows your dataset back to you through the FiftyOne interface so you can curate the images manually. That curation step is the point of the tool rather than an afterthought, since a training run inherits every flaw in the set you give it. Tag generation comes from the Waifu Diffusion 1.4 Tagger model, caption generation from the BLIP model, and you can edit hundreds of text files at once to add, remove or replace tags inside them. Everything works inside your Google Drive by default, and the notebook connects easily to the Lora Trainer that follows it.

XL training with presets chosen for you

The XL Lora Trainer trains both LoRA and LoCon for Stable Diffusion XL and includes several model options for anime. Installation and training start from a single click, and it can run on several Colab configurations including T4, which is the free tier, and A100. The parameter story is the most interesting design choice here. The trainer offers most of the parameters while setting useful values behind the scenes to keep it simple, which is a deliberate trade: you can adjust a lot, but the defaults are chosen to produce a sensible result rather than to be a faithful mirror of every upstream flag. You can also define multiple folders for training, and it works inside your Google Drive by default and connects easily to the Dataset Maker. The legacy XL trainer is kept in the repository as a separate notebook, presumably for users whose workflow predates the current one.

Getting to the notebooks without a command line

There is no installation step in the usual sense, and that is the point. The project distributes its work as Colab notebooks linked straight from the project page, with an Open in Colab badge for each one, so the entry point is a browser and a Google account. The Dataset Maker, the Lora Trainer, the XL Lora Trainer and the legacy XL trainer are each one click from the table at the top of the page, and the Spanish versions are reached from the same table. Two supporting links sit alongside them: a Lora making guide on Civitai, and a Stable Diffusion guide hosted on Hugging Face with English and Spanish indexes. Support is offered through a public Discord server, and the author links a coffee funding page. No package manager, no virtual environment and no driver installation appears anywhere in the documentation, because Colab provides the runtime.

What this approach gives up against a local setup

The comparison worth making is against running the upstream kohya-ss scripts on a machine you control, because that is what these notebooks wrap. A local setup gives you a pinned dependency set, a GPU you choose, a script you can re-run from a config file, and outputs you own outright. The notebook route gives you a working environment immediately, which for a first LoRA is worth a great deal. What it costs is control and reproducibility. A notebook is a sequence of cells that ran once in a session, not a versioned script, so turning a successful run into a repeatable pipeline means rewriting the steps by hand. The training data and outputs also live in a Google Drive folder that belongs to a Google account rather than to your infrastructure. None of this makes the notebooks wrong; it defines what kind of task they suit.

Licence, maintenance and the absence of releases

The project is released under GPL-3.0, which is a copyleft licence that obliges you to keep distributing modified versions under the same terms. That matters more here than for most small utilities, because a trained LoRA model produced with these notebooks may fall under the licence terms of the models and scripts involved, and the guidance on that belongs to the model authors and to your own legal review rather than to this repository. On maintenance, the repository is not archived and the last push was on 2026-09-26, so the notebooks are being kept in step with upstream changes. There are no GitHub releases, which means no tags to pin and no changelog to diff; you take whatever the notebook files contain at the moment you open them, and updating means pulling the repository again. Given that the notebooks track moving upstream toolkits, that lack of versioning is the single biggest operational risk of the approach.

Editorial conclusion

kohya-colab fits an artist or hobbyist who wants to train a Stable Diffusion LoRA on Colab infrastructure rather than own a GPU, and who is comfortable working through a notebook cell by cell. It is the wrong choice if you need a reproducible pipeline, because a notebook held in someone else's runtime is not a versioned script you can re-run unattended, and the project publishes no releases to pin. Before you start, check that your dataset is curated and correctly tagged, since the trainer will happily fit whatever images and captions you hand it, and decide in advance how you will move the resulting LoRA out of the Drive folder when the session ends.

Frequently asked questions

Official sources

1. hollowstrawberry/kohya-colab on GitHub
2. Issues
3. License: GPL-3.0
4. README