MinusZone - KohyaSSTrain(lora)

This is the node to use. MZ_KohyaSSLoraTrain (display name "KohyaSSTrain(lora)") is the pack's current, merged LoRA training node: it takes the workspace, the template choice, the base checkpoint, and the training basics all in one place, then runs the kohya training the way the old two-step chain (MZ_KohyaSSUseConfig → MZ_KohyaSSTrain) did. If you're starting a MinusZone workflow from scratch today, this is the one the graph is built around.

The workflow shape is: MZ_KohyaSSInitWorkspace → MZ_KohyaSSDatasetConfig → this node, with MZ_KohyaSSAdvConfig optionally feeding advanced_config. Everything that matters for the run lives on these three.

How it works

Like the old trainer, it builds a full sd-scripts config from a template JSON in the pack's configs/kohya_ss_lora/ folder (lora_sd1_5, lora_sdxl, lora_hunyuan1_1, lora_hunyuan1_2, controlnet_sd1_5), then runs training in-process. The difference from the old chain: the template, checkpoint, steps, and learning rate are all set right here on the node, and the optional advanced_config from MZ_KohyaSSAdvConfig overrides template defaults. No separate config-builder step.

It also carries the resume trick from the old trainer: base_lora defaults to latest, which means "continue from the newest .safetensors in this workspace's output folder." Combined with save_every_n_epochs, that's your stop-and-resume story - crash, fix the env, re-queue, pick up where the last save was.

The inputs you'll actually set

• train_config_template - your model family. This matters more than anything else on the node. lora_sd1_5 for SD 1.5, lora_sdxl for SDXL, the hunyuan ones for Hunyuan DiT via this path.
• ckpt_name - the base checkpoint from your ComfyUI checkpoints folder.
• max_train_epochs (default 100) / save_every_n_epochs (default 10) - train-to-save cadence. Drop save_every_n_epochs to 2–3 so you have checkpoints to pick from; the last epoch is rarely the best.
• learning_rate - a string like "1e-5" (sd-scripts convention). Don't type 0.00001 and wonder why it's treated as text - it still works, but kohya parses it as a string either way.
• base_lora - latest to resume, empty to start clean, or any scanned LoRA.
• sample_generate + sample_prompt - renders a sample at each save and shows it in the UI. Keep it on; it's your early-warning system for overfitting.
• advanced_config (optional) - from MZ_KohyaSSAdvConfig. caption_completed_flag (optional) is an ordering signal - wire the output of a captioning node here to force sequencing.

Install & troubleshooting

cd ComfyUI/custom_nodes
git clone https://github.com/MinusZoneAI/ComfyUI-TrainTools-MZ
# restart ComfyUI

Or Manager → search ComfyUI-TrainTools-MZ.

First-run gotchas: the pack clones sd-scripts at a pinned commit on the InitWorkspace step (not here), and the first train triggers auto-pip of helper deps. If you chose an 8-bit optimizer upstream, install bitsandbytes yourself or the run dies at startup. And the dataset gotcha is upstream too - this node reads workspace_images_dir from the dataset step, and if that string never got wired, training starts against an empty dataset. The base_lora = latest behavior also means a fresh run over an old workspace silently continues the old LoRA - use empty (or a new lora_name) when you want a clean slate.