RunningHub_ZImageI2L_Loader

The Loader is the front door to the whole image-to-LoRA pack, and it's the node with the most unglamorous job in the chain: you drag it in, change nothing, and it produces the ZImageI2LPipeline socket that the LoraGenerator refuses to work without. There are no inputs and no settings - it's a one-trick node that says "load everything" and does exactly that, loudly.

What it actually loads is where the interesting stuff hides. Instead of dropping a checkpoint into ComfyUI's normal model slots, it builds a full DiffSynth-Studio ZImagePipeline from six separate model bundles: the Z-Image base transformer, the Z-Image-Turbo text encoder, VAE and tokenizer, two image encoders (SigLIP2-G384 and DINOv3-7B), and the Z-Image-i2L model itself. That's tens of gigabytes across six safetensors. If this is your first run, the node will sit there for a while downloading all of it from ModelScope (not HuggingFace) into ~/.cache/modelscope/hub/ - set the MODELSCOPE_CACHE env var if you want it somewhere else. A long "hang" on first use is usually just the download; watch the console for progress.

The load itself is clever-ish but heavy. Everything runs in bfloat16 with a disk-offload config, so the big transformers shuffle between CPU and GPU instead of all living in VRAM at once. Two things worth knowing before you click "Run":

• It unloads your current models first. The source calls model_management.unload_all_models() before building the pipeline, then garbage-collects. If you had a checkpoint loaded for a previous workflow, it gets evicted. Wire this node at the start of your graph, before you load the Z-Image checkpoint you'll actually generate with.
• It's VRAM-hungry regardless. The README asks for 24GB+ and says it was tested on a 4090. The offload may let it squeeze onto less, but don't count on a 12GB card coasting through this. This is the heavyweight side of Z-Image - a very different vibe from the 30-second-per-image Turbo inference the model family is known for.

Install is the standard two-step - via ComfyUI Manager searching "ComfyUI_RH_ZImageI2L", or:

cd ComfyUI/custom_nodes
git clone https://github.com/HM-RunningHub/ComfyUI_RH_ZImageI2L.git
cd ComfyUI_RH_ZImageI2L
pip install -r requirements.txt

then restart ComfyUI. The requirements file is small, but diffsynth is the real dependency - it drags in the whole DiffSynth-Studio stack that the Z-Image pipeline lives in.

If you're troubleshooting: a black screen or OOM after the Loader usually means the six models didn't all fit, and the fix is less about settings and more about VRAM headroom. And if a run seems to "forget" your previously loaded model, that's not a bug - the Loader deliberately cleared it. It's a crude but effective way to make room for a very big pipeline.