ZILORA
======
ZILORA is a simple tool (under 650 lines) for FLUX.2 Klein 4B or Z-Image Base
LoRA merging and extraction. It works with 32 GB RAM (or 16 GB RAM and 16 GB
swap). During the LoRA extract, CUDA and around 2 GB VRAM is used. The BF16 and
CUDA are intentionally fixed as that's the hardware I use at the moment.
  The tool is a command-line tool for Linux and Python environments.

USAGE
-----
The tool supports dump, extract and merge operations. Run it with
"python zilora.py -h" to get the internal help text which is the main
documentation.

LoRA FORMATS
------------
The LoRA scene is a mess! This tool supports some LoRAs I've encountered, with
the following Z-Image key variations:
  adaLN_modulation.0
  attention.out
  attention.qkv
  attention.to_out.0
  attention.to_q[qkv]
  feed_forward.w[123]
It assumes "lora_A" and "lora_B", and "alpha" is used if present but not
written. FLUX.2 Klein 4B is simpler as there seems to be only a single format
around.
  Only "officially" supported formats are ones produced by musubi-tuner. Other
formats mainly work just by accident. The output format is intended to be
compatible with both musubi-tuner and InvokeAI.

CREDITS
-------
The original merging and extraction code was forked from
<https://github.com/GraftingRayman/ComfyUI-ZImage-Bake-Extract>.

LICENSE
-------
Apache-2.0 (the original license of ComfyUI-ZImage-Bake-Extract)