Z-Image Base LoRA · Text to Image

Z-Image Base is the interesting half of Alibaba's Z-Image family: 6.15B parameters, Apache 2.0, uncensored, and - unlike the 8-step Turbo everyone actually generates with - the one you train LoRAs on. Training a LoRA on Base means it reaches both Base and Turbo, which is why Base matters even if you'd never sample with it. This node runs Base as a hosted text-to-image endpoint with up to three of your own LoRAs attached, for $0.0144 an image.

Why this node exists

Local Z-Image on a 3060 takes about half a minute a picture, and a LoRA test sweep of sixty prompts is therefore a thirty-minute coffee break. Here it's under a dollar and you get flat per-image pricing, so the sensible loop is: iterate the dataset and the training locally, then validate the LoRA against a real prompt set on SpicyAPI. You also get Base's seed diversity and its ability to listen to a negative-ish sentence, which the distilled Turbo is weaker at.

The other honest use: you want a LoRA you found on Civitai or Hugging Face without downloading a 12B-class stack and a matching workflow. Paste the .safetensors URL, set a strength, done.

The inputs

• prompt - write it in sentences: subject, clothing, place, light. If your LoRA has a trigger word, put it first, every time. Missing the trigger is the number one reason a LoRA looks like it did nothing.
• aspect_ratio - 1:1 by default, with nine ratios available in total.
• output_format - jpeg (default), png or webp.
• seed - the same seed and settings give the same image, so this is your variation control.
• Optional loras - a SPICY_LORAS socket fed by a SpicyAPI LoRA node, which takes a direct .safetensors link and a strength; you can chain several LoRA nodes to stack them. Up to three per request, and the link must open without a login - a Google Drive or Civitai page that wants a session won't work, a raw Hugging Face file URL will.

Outputs: images (IMAGE - Save Image, or a preview/upscale chain) and task id (STRING, for logs and support).

Install

cd ComfyUI/custom_nodes
git clone https://github.com/spicyapi-ai/comfyui-spicyapi

Or ComfyUI Manager → search SpicyAPI. Restart ComfyUI, then Settings > SpicyAPI → paste a key from spicyapi.ai/console/keys, and restart again so the node list is rebuilt from your account's catalogue. Remote box? Set SPICY_API_KEY. There are no weights to download and nothing to pip install; the pack's dependency list is empty and it uses only what ComfyUI already ships. It wants Python 3.10+ and the V3 node API (tested with ComfyUI 0.39).

Gotchas

Stacking is the one to watch. Z-Image tolerates one LoRA well and gets wobbly with several, which is why the KB's own guidance is to avoid piling them up - the three-LoRA limit on this node is a maximum, not a target. Keep strengths low when you do stack. Likewise, don't expect Base to be fast: it's the non-distilled model, and a cent and a half buys you quality, not the "photoreal frame in roughly a second" feel of the Turbo endpoints.

Cost control is simple at this price, but set Max cost per run anyway if you plan to queue big sweeps, and leave the seed on fixed while comparing prompts so a re-queue reuses the cached image instead of re-billing. Failures are usually one of: No SpicyAPI key is set, 40310 (verify your email), 40004 (no deployment serves that combination - change the named option), or a LoRA that refused to load because its URL needs a login.