Z-Image i2L v2 — Save LoRA

Extract LoRA hands you a LoRA living in memory. This node is what makes it a file: it writes the .safetensors into ComfyUI's models/loras folder, where any normal Z-Image LoRA workflow can load it later. That turns "the style of these four images" into a reusable asset you can keep, share, or drop into a completely different graph tomorrow.

Two inputs, both trivial: lora (from Extract LoRA) and filename (default zimage_i2l_v2_style - it appends .safetensors if you forget). The only output is the filename string, confirming what got written. It's an output node, so it's the end of a branch - nothing plugs into its other side.

The workflow: Loader → Extract LoRA → Save LoRA is the README's first usage pattern, and it's the one that makes sense when you want the style again on a different prompt, in a different workflow, or as a file on disk. After saving, load it with any stock LoRA loader in a regular Z-Image graph.

Here's the honest caveat, straight from the README: this saved-file path is the portable version, not the best-quality version. The i2L paper's trick is asymmetric CFG - a style LoRA runs on the positive branch, and a neutral-gray-image LoRA runs on the negative branch. The Generate node keeps that automatically. A saved LoRA loaded through ComfyUI's stock loader gets applied to both CFG branches the same way, which isn't how the i2L style was tuned. So: saved file for portability and reuse, Generate for the best single image. If you're only ever going to generate once, skip Save entirely and use Generate.

Troubleshooting is thin because it's a dumb node. Reuse a filename and it silently overwrites - put something identifying in the name. It writes to the first "loras" folder in ComfyUI's folder paths (usually models/loras), so if you've pointed loras elsewhere in extra_model_paths.yaml, check that location. And note the pack's own Sample node is the better place to use a fresh in-memory LoRA - you only need this node when you want the file.

Install

Install the pack via ComfyUI Manager (search ComfyUI_ZImageI2L_v2) or git clone https://github.com/jzhang-POP/ComfyUI_ZImageI2L_v2 into ComfyUI/custom_nodes/. The dependency that bites is DiffSynth-Studio, which must be installed from git because v2 needs its Diffusion Templates API (diffsynth.diffusion.template.TemplatePipeline):

git clone https://github.com/modelscope/DiffSynth-Studio.git
cd DiffSynth-Studio && pip install -e . && cd ..
pip install -r ComfyUI/custom_nodes/ComfyUI_ZImageI2L_v2/requirements.txt

Restart ComfyUI, and let the Loader complete its one-time model download (tens of GB from ModelScope) before you run the extract-and-save chain.