kohya-ss — Kohya LoRA / fine-tune trainer
The standard toolkit for training Stable Diffusion LoRAs, LyCORIS, DreamBooth, and full fine-tunes. Bundled with the Kohya GUI.
Image tag
docker.io/manvarharsh/kohya-ss:cuda12
What’s in this image
- Base:
nvidia/cuda:12.4.1-cudnn-devel-ubuntu22.04 - Python 3.10 (conda)
- PyTorch with CUDA 12
- bmaltais/kohya_ss (with submodules)
- bitsandbytes, accelerate, xformers
- OpenSSH server
Default ports
| Port | Service |
|---|---|
| 22 | SSH |
| 7860 | Kohya GUI |
Use cases
- Training SD 1.5 / SDXL LoRAs on character / style datasets
- DreamBooth fine-tunes
- Textual inversion embeddings
- LyCORIS variants (LoCon, LoHa, LoKr)
- Batch dataset captioning + bucketed training
Environment variables
| Variable | Description |
|---|---|
ENABLE_SSH | Enable SSH server |
ENABLE_KOHYA | Start the GUI on port 7860 |
KOHYA_EXTRA_ARGS | Extra CLI args for kohya_gui.py |
SSH_PUBLIC_KEY | Public key for SSH |
Persistence
Mount at /data. Put training images under /data/datasets/<name>/, base checkpoints under /data/models/, and outputs under /data/output/loras/.