FL Kohya Adv Config

If FL Kohya Easy Train is the "I just want it to work" path, FL Kohya Adv Config is the "I know exactly what I want" path. It's a form over every training flag Kohya's sd-scripts accepts, and it's where the FL-Trainer manual chain gets its personality. You set your values here, the node hands a config object to FL Kohya Train, and your choices get merged into the training run.

It only exists because of what Kohya is: a CLI with dozens of flags and a config convention the whole ecosystem copied. This node is a GUI for that vocabulary - network_dim, network_alpha, lr_scheduler, optimizer_type. If you can read a Kohya config dump, you already know this node; if you can't, this is a friendlier way to learn the words.

How it works

The node doesn't train anything itself. It collects your settings and emits an advanced_config object. When FL Kohya Train runs, each value is layered over the base config template: enable/disable become booleans, numbers become strings, and - the useful trick - an empty string means "leave the template's value alone." So unet_lr left blank simply inherits your main learning rate, and you only override the things you actually care about.

Inputs that matter

• network_dim / network_alpha - rank and scaling, the two numbers every Kohya guide obsesses over. Defaults are 16/8, the classic "alpha at half rank" convention. Rank 32 with alpha 16 is the SDXL character-work sweet spot if you want more capacity.
• network_module - networks.lora (default), plus networks.dylora and networks.oft for the exotic variants.
• network_train_unet_only - on by default, meaning the text encoder is left frozen. For SDXL that's a big VRAM saver and most people never flip it; leave it on.
• optimizer_type - 18 choices from plain AdamW up through AdamW8bit, PagedAdamW, Lion, the DAdapt* family and beyond. AdamW8bit is the safe pick; Adafactor trades quality for lower VRAM.
• lr_scheduler - cosine (default), cosine_with_restarts, constant, polynomial, and friends. The EasyTrain node quietly uses constant; you get to choose.
• unet_lr / text_encoder_lr - separate LRs for the two halves. Empty inherits; setting just one is a common trick to let the U-Net learn while the encoder coasts.
• shuffle_caption - shuffles your tags each epoch, a real quality boost for tag-style datasets.
• cache_latents + cache_latents_to_disk - precompute latents once and reuse across epochs. Both default on; this is the single biggest speed lever for multi-epoch runs.
• gradient_checkpointing, gradient_accumulation_steps, mixed_precision, xformers/sdpa, fp8_base, noise_offset, save_precision, lr_warmup_steps - the VRAM-and-precision dials. You'll grow into these.

The output is advanced_config, which feeds directly into FL Kohya Train's advanced_config input.

Install

Same as the whole pack - Manager search "FL-Trainer", or git clone https://github.com/filliptm/ComfyUI_FL-Trainer into custom_nodes and restart.

Gotchas

• Defaults match the SD 1.5 template, not SDXL. The node ships network_train_unet_only on, rank 16/8, fp16. Fine for both, but check resolution upstream and your VRAM before blaming a bad run on settings.
• Empty-string fields are "unset," so you can't force an explicit empty value into a flag - a mild trap if you're porting a config that genuinely passes an empty string.
• The manual chain is SD 1.5/SDXL only (pinned sd-scripts), same as the rest of this pack.
• If you just want a working LoRA, you don't need this node at all - EasyTrain sets all of this for you. Reach for Adv Config when EasyTrain's hidden choices aren't good enough and you know why.

The one caveat: a bad dataset beats perfect settings every time. But for the second run, when you're chasing a real result, this is where the power is.