Z-Image API Config

The name is a lie, in the good way: this node doesn't call any API and needs no key. It's the single configuration hub for the whole pack - the thing that decides which LLM powers your Z-Image prompt enhancement, and how that LLM is reached. Every graph that uses this pack starts here, because both Z-Image Prompt Enhancer and Z-Image Integrated KSampler take a config input that only this node produces. Despite the "API Config" label, it's really a connection builder: pick a backend, fill in its details, and it hands you a config object the rest of the pack can use.

Three providers, one dropdown:

OpenRouter (cloud, easiest)

provider: openrouter with a model ID like the default qwen/qwen3-235b-a22b:free. Paste an API key from openrouter.ai/keys into api_key. The free tier genuinely works - the author's own demo runs were done on the free Qwen model. Want to feed an image into the enhancer? Use a vision-capable model here, or the image input will be rejected.

Local (your own server, no key)

provider: local points at any OpenAI-compatible /v1/chat/completions server. The local_endpoint field defaults to http://localhost:11434/v1 (Ollama), with LM Studio on :1234, vLLM on :8000, and text-generation-webui on :5000 as the common alternatives. model is the model name your server exposes - e.g. qwen2.5:7b. This is the route that costs nothing per run and keeps everything on your machine; several users run a small Qwen in LM Studio on a second GPU just to feed this node.

Direct (HuggingFace, zero servers)

provider: direct downloads and loads the model itself. model is a HuggingFace repo ID, downloaded into ComfyUI/models/LLM/Z-Image/ and cached. This is where the heavy settings live:

• quantization - 4bit (default), 8bit, or none. 4-bit saves the most VRAM; none (fp16) is best quality. The tooltip's honest framing.
• device - auto, cuda, cpu, or mps.
• llm_path - an optional local model folder; if set, the model name is treated as a folder name inside it.
• auto_download_fallback - on: try the manual path, fall back to downloading if not found. Off (default): strict, fails if the folder is missing.

The Direct route is the only one with real dependencies: pip install bitsandbytes huggingface-hub (hf_transfer optional for faster downloads). And it's the only route where you might OOM - dropping to 4-bit or switching to a smaller model is the fix.

Output and gotchas

Single output: config (type ZIMAGE_CONFIG), feeding the Prompt Enhancer or Integrated KSampler. A few things that'll bite: an empty model errors immediately; OpenRouter with no key logs a warning and then fails at request time; and a text-only Direct model will reject image input - pick a vision-language model for that. If your local server is misconfigured, the node will still happily build a config and only fail when the enhancer actually calls the endpoint, so keep the Z-Image Options debug_mode in mind when troubleshooting.

Install the pack once via ComfyUI Manager (search "Comfyui-Z-Image-Utilities") or cd ComfyUI/custom_nodes && git clone https://github.com/Koko-boya/ComfyUI-Z-Image-Utilities.git, restart, and this node is your front door.