NAI Model Loader

NAI Model Loader looks like every other model loader in ComfyUI and is nothing like it. A normal Load Checkpoint node reads gigabytes of weights off your disk. This one has a single dropdown, four choices, and downloads nothing - because the "model" here is a pointer to a model running on NovelAI's servers, and your job is just to pick which one the Sampler calls. That distinction is the whole node, and missing it leads to the most common confusion with this pack.

The four choices

The model enum is the entire interface:

• V5 Full (default) - NovelAI Diffusion V5, full model.
• V5 Curated - the curated V5 route.
• V4.5 Full - NovelAI Diffusion V4.5.
• V4.5 Curated

Under the hood each label maps to a model id (nai-diffusion-5-full, nai-diffusion-4-5-curated, and so on), and the node's single output, also called model, is that id wrapped in the pack's NAI_MODEL type. Plug it into NAI Sampler's model input and you're done. The source is honest about one oddity: V5 Curated inpainting is routed through V4.5 Curated inpainting on NovelAI's side, so that mode keeps V4.5's six-character limit.

The V5 reference gotcha

Here's the setting that bites people. V5 silently ignores connected Reference Resources. Wire up your Vibe or Precise Reference cards, then switch the Model Loader from V4.5 to V5, and the references just... stop mattering. The README is explicit that this matches the intended V5 workflow, but "silently" is the operative word - there's no warning node. If your reference-style isn't showing up, check which model you picked before you blame the reference cards. V4.5 uses the configured Vibe or Precise Reference cards; V5 doesn't.

Install

Identical to the rest of the pack: ComfyUI Manager (search 4A NovelAI) or

cd ComfyUI/custom_nodes
git clone https://github.com/tsukino4a/ComfyUI-4A-NovelAI.git ComfyUI-4A-NovelAI
cd ComfyUI-4A-NovelAI
python install.py

then restart ComfyUI. Because there are no local weights, there's also no VRAM cost from this node and no giant models/ folder to manage - the tradeoff being that every generation costs Anlas and your prompts go to NovelAI's servers. If you expected a local checkpoint you came to the wrong pack; this one is purely a remote model selector.