LoRA Folder Batch

Every LoRA training run ends the same way: twenty checkpoint files from twenty epochs, and the best one is never the last one. LoRA Folder Batch is built for exactly that moment. Point it at a folder of .safetensors LoRAs, tell it the strengths you want to test, and it generates a batch of LoRA-parameter sets you can sweep through - epoch by epoch, strength by strength - without hand-plugging each file into a LoRA loader.

How it works

It scans a folder for .safetensors files (relative to your models/loras directory, or an absolute path), then applies filters and sorts before generating params. The natural-sorting detail is the one that shows someone actually trained models: it sorts by epoch number, so epoch_004 comes before epoch_020 - alphabetical sort would put epoch_020 first and your sweep would be garbage. You can also sort alphabetically, newest, or oldest.

Strengths are a flexible string: "1.0" for a single value, "0.5,0.75,1.0" for a list, or "0.5...1.0+0.1" for a range sweep. batch_mode controls the combinatorics: sequential pairs each LoRA with one strength, combinatorial tries every strength on every LoRA (that's the exhaustive sweep, and it multiplies your run count fast). Regex include_pattern / exclude_pattern fields let you slice the folder down to, say, the character LoRAs and skip the style ones.

Inputs and outputs

The ones that matter:

• folder_path - relative to models/loras (default .), or absolute.
• strength - single value, comma list, or start...end+step range.
• batch_mode - sequential or combinatorial.
• max_loras - cap, default 50, to stop a giant folder from flooding your UI.
• auto_batch / batch_size / batch_index - for chunking huge sets into groups of 25 (default) so the UI doesn't choke.

Outputs: lora_params (the batch, meant for the pack's Flux Sampler Params), lora_list (a readable list), and lora_count (how many made it through).

Where it fits

This is the "find the best epoch" tool, and the training lore backs it up: intermediate epochs are routinely better than the final one, so testing all of them is not optional paranoia. It's also the strength-sweep workhorse - you'll often find a LoRA that "doesn't work" just wants 0.4 instead of 1.0. It's part of the pack's xyz-helpers family (descended from cubiq's ComfyUI_essentials), so it's designed to feed the pack's parameter-testing plumbing, not a stock LoRA loader.

Installing it

ComfyUI Manager → "ComfyUI-KikoTools" → Install → restart, under ComfyAssets → xyz-helpers. Manual:

cd ComfyUI/custom_nodes
git clone https://github.com/ComfyAssets/ComfyUI-KikoTools.git
cd ComfyUI-KikoTools
pip install -r requirements.txt

Restart ComfyUI.

Gotchas

Watch the run-count explosion: combinatorial mode across 20 epochs × 3 strengths is 60 generations - plan accordingly or use the max_loras / batch_index chunking to eat it in pieces. If you also have the original ComfyUI_essentials installed, the shared lineage can surface duplicate node names; ComfyUI Manager usually flags it. And remember the folder path is relative to models/loras by default - put your training output there (or give an absolute path), or the scan comes back empty.