Training an AI on 22 Photos of My Cat: LoRA Model Generates Custom Scenes
Train a LoRA model with your own images by converting HEIC to JPG, resizing to 512px, generating captions, and running Kohya_ss with 1100 steps.
Train your own LoRA model using the described pipeline to generate custom AI art.
Summary
Training a LoRA model on 22 photos of my cat required converting 10 HEIC files to JPG, resizing each image to 512px on the short side, and generating captions that include the trigger token “ohwx”. The captions were drafted by Claude and reviewed manually, producing concise descriptions such as “ohwx cat, walking on a metal kitchen counter, side profile, indoor kitchen with spice bottles and shelves in the background”. The dataset was fed into Kohya_ss, running accelerate launch train_network.py with a 1100‑step training session that finished in 13 minutes and 3 seconds on a DGX Spark. The resulting LoRA model, when prompted with “ohwx cat”, produced the cat in a variety of scenes—from a chef in a kitchen to an astronaut in a helmet—demonstrating that a small, curated dataset can generate high‑quality custom AI art. The experiment shows that building your own AI from your own data is more accessible than previously assumed. The key takeaway is that LoRA training is straightforward and yields impressive results with minimal data.
Key changes
- 22 cat photos converted from HEIC to JPG and resized to 512px on the short side.
- Each image received a caption with the trigger token “ohwx” to identify the cat in prompts.
- LoRA training performed with Kohya_ss, running 1100 steps in 13 minutes on a DGX Spark.
- The resulting LoRA model produced the cat in diverse scenes such as a chef, balcony, forest, and astronaut.
- The ohwx token became a valid prompt token, enabling the model to generate the cat when used.
- The training pipeline demonstrated that small, curated datasets can produce high‑quality custom AI art.