NAI 模型加载器

Every GENYTOOLS workflow starts here, and you should know what it isn't before you worry about what it is. This is not a checkpoint loader in the ComfyUI sense - there are no .safetensors files, no VRAM, no download progress bar. The model lives on NovelAI's servers, and this node is a remote control for picking which one the API uses. It's an API-wrapper pack, the whole "the model isn't on your machine" category: you need a NovelAI account with a Persistent API Token, and every generation you run spends Anlas. The node itself is free and instant.

What it does is wonderfully simple. One dropdown, eight choices: V5 精选版 (Curated) and V5 完整版 (Full), the same pair for V4.5 and V4, plus 动漫 V3 and 兽人 V3. Pick one and it hands you an NAI_PIPE on one output and a balance readout on the other. The pipe is this pack's version of the context-bus idea - model, plus anything later nodes bolt on (Vibes, character regions, precise references), travels down a single wire into the sampler. You don't plumb a model, CLIP and VAE separately because there's nothing local to plumb.

The dropdown is where the real knowledge lives, because the models are not interchangeable. NovelAI's capability matrix is genuinely per-model, and the pack's docs spell it out:

• V5 Full / Curated - up to 32 character regions, transparent backgrounds, Max✨ Enhance. No Vibe transfer and no precise reference yet; NovelAI hasn't opened those to V5.
• V4.5 Full / Curated - up to 6 character regions, plus Vibe and precise reference.
• V4 Full / Curated - 6 character regions, Vibe, no precise reference.
• Anime V3 / Furry V3 - no character regions, but SMEA and dynamic thresholding (which the newer models dropped).

That's why the wrong pick shows up as a hard error, not a bad image: the pack checks the combo before it spends your money and raises a Chinese-language message instead of silently dropping the unsupported feature. That's a genuinely good design - silent param-dropping is how API packs quietly waste your Anlas.

The 积分与订阅 (balance and subscription) output is a string with your account summary - Opus subscribers get a free-generation allowance on standard-res V5 runs at up to 28 steps, and the loader will show your estimated remaining free images. Handy, but it's a readout, not something you wire into anything.

Install and token

This is the one install that matters for the whole pack:

cd ComfyUI/custom_nodes
git clone https://github.com/XTOGENY/ComfyUI-NovelAI-GENYTOOLS.git
pip install -r ComfyUI-NovelAI-GENYTOOLS/requirements.txt

or just search "ComfyUI-NovelAI-GENYTOOLS" in ComfyUI Manager. Dependencies are light - requests, Pillow, numpy, msgpack - because nothing runs locally. Then open ComfyUI settings and paste your Persistent API Token into NovelAI 2.0：Persistent API Token. The pack deliberately keeps the token out of workflows, image metadata and logs, which is more than most API packs bother with.

The gotchas

• "尚未配置 NovelAI Persistent API Token" on first run just means you skipped the settings step. Go set it.
• Switching models mid-pipeline can invalidate cached Vibes - the cache key includes the model, so the encoder will want to re-encode (and re-charge) after a swap.
• Vibe nodes will refuse to run at all on V5. That's the current model limitation talking, not a bug.

This pack is brand new and largely unreviewed as of writing, so don't expect a big community trail yet - read the README's examples before you let anything loose on paid requests.