MinusZone - KohyaSSAdvConfig

If MZ_KohyaSSInitWorkspace is the pack's front door, MZ_KohyaSSAdvConfig is the control panel with the confusingly many switches. It's a single node that collects the sd-scripts training flags you'd normally sprinkle through a kohya config file, and hands them forward as one advanced_config object. You don't need to touch most of them - the pack's templates already have sane defaults - but when you outgrow the template, this is where you grow.

Everything here maps 1:1 to a kohya flag, which is the same vocabulary the KB's LoRA training doc spends pages on. That's both the strength and the danger: the node exposes the raw knobs with no presets, and the kohya-ss panel's verdict is that this is exactly why beginners bounce off. Start with the template, tweak from there.

The inputs that actually matter

• network_dim / network_alpha (defaults 16 / 8) - the rank and the alpha. The alpha-at-half-rank convention (alpha ≈ dim/2) is Kohya's house style, and it's what "rank 16, alpha 8" means when you read a config dump. Raise network_dim when a small rank underfits a style, not when your character looks off.
• network_module - networks.lora (default), networks.dylora (dynamic, more expressive, fiddlier), or networks.oft. Stay on lora for your first few runs.
• mixed_precision - fp16 default, bf16 on cards that support it, no if you have VRAM to burn and want to fight for precision.
• cache_latents + cache_latents_to_disk (both default enable) - precompute the latents once instead of every epoch. The single biggest speed lever in a modern config; leave both on.
• optimizer_type - AdamW default, with the whole 8-bit / Paged / Lion / DAdapt family in the dropdown. The KB notes 8-bit optimizers are where installs usually break: AdamW8bit needs bitsandbytes, which is the most dependency-prone package in the kohya stack. If you pick one and training dies at startup with a bnb error, that's why.
• lr_scheduler / lr_warmup_steps - cosine with a small warmup is the boring, correct choice.
• network_train_unet_only (default enable) - train only the UNet, not the text encoder. Common on SD-family LoRAs, and it saves memory.
• noise_offset (0.1) - a small value helps contrast; the "0.1" everyone copies.
• xformers / sdpa - the attention backends. Enable one, not both; sdpa is the modern default, xformers is the legacy fallback.

How it works & what it outputs

The node just packages the dict - the actual training flags get merged into the final config when MZ_KohyaSSTrain or MZ_KohyaSSLoraTrain runs. Output is advanced_config (type MZ_TT_SS_AdvConfig). In the old chain you wire it into MZ_KohyaSSUseConfig's save_advanced_config; in the new chain straight into MZ_KohyaSSLoraTrain. Either way, values here override the template.

Install & troubleshooting

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

Or ComfyUI Manager → ComfyUI-TrainTools-MZ. No extra deps for the node itself.

Where people get burned: flipping xformers and sdpa both on (some environments tolerate it, some don't), picking a save_precision of float and wondering why the output is twice the size, and choosing an 8-bit optimizer without bitsandbytes installed. And the eternal one - a higher network_dim is not free quality. If you're tuning because results are bad, change one thing at a time, and remember the KB's blunt line: dataset curation beats every knob.