Trainers (kohya sd-scripts, OneTrainer)

Forge writes training configs for kohya sd-scripts, OneTrainer and diffusers, and can run kohya itself in its -train image.

Used by

Package	What for	Reads
argus-forge (opens in new tab)	`kohya`: the forged `train.sh` runs from `SD_SCRIPTS_DIR` (preset to `/opt/sd-scripts` in the `-train` image).	`SD_SCRIPTS_DIR`, `argus-forge run --env`
argus-forge (opens in new tab)	`onetrainer`: config files only, loaded in OneTrainer's own UI. `diffusers`: points at a training script.	`DIFFUSERS_SCRIPT`

Official guides

Minimal setup

The quickest way is forge's training image, which ships kohya sd-scripts at /opt/sd-scripts:

docker run --rm --init --gpus all --shm-size=8g \
  -e ARGUS_FORGE_READONLY=0 \
  -p 127.0.0.1:8103:8103 -v <export-dir>:/data/out \
  -v hf-cache:/root/.cache/huggingface \
  ghcr.io/smk762/argus-forge:latest-train

An armed POST /run executes a forged script with no authentication, so treat reaching that port as shell access on the host. Keep it on loopback or a trusted network.

To run the forged train.sh on your own machine instead, clone kohya-ss/sd-scripts, install its requirements, and set SD_SCRIPTS_DIR to the checkout. The configs refer to container paths (/data/out, /data/images), so swap in your own OUTPUT_DIR and DATASET_DIR, or set FORGE_PATH_MAP.

OneTrainer configs are loaded in OneTrainer's own UI. Forge doesn't run them.