GPT-5.4

ComfyUI isn't just for images anymore, and this node is the argument: a GPT-5.4 chat model on your canvas, able to read the images in your workflow and answer in plain text. Prompt-rewriting, style-guide generation, upscaling advice that's aware of what's on screen, structured reasoning - all of it without leaving the graph or installing a single model file.

How it works

A textInference node that calls OpenAI's GPT-5.4 through the Runware cloud API. The required input is messages, and it's a RUNWARE_MESSAGES socket - you don't type a chat log into a widget, you build it with the Runware Messages builder node (look in Runware/Params). That's the pack's builder pattern: stackable inputs keep the model node clean.

What you'll actually configure:

• settings.systemPrompt - the system-level instruction steering the whole exchange. This is where you set the persona ("you are a prompt engineer") or the constraint ("reply in valid JSON").
• settings.temperature (default 1, range 0–2) and settings.topP (default 0.95) - the classic creativity dials. Lower temperature for deterministic rewrites, higher for ideation.
• settings.thinkingLevel - reasoning depth: (default), low, medium, high, xhigh. Crank it for hard problems, leave it low for trivial routing.
• settings.maxTokens (default 4096, up to 128k) - response length cap.
• images (IMAGE) - optional vision: attach any image in your workflow and the model can see it. This is the sleeper feature - ask the model about the image you just generated.
• toolChoice / toolChoice.type (auto/any/tool/none) with toolChoice.name for forcing a specific tool, plus advanced_json where you drop in your JSON tools definition. That's the function-calling escape hatch.
• seed, numberResults (up to 4 variants), includeUsage (token stats in the response).

Output is a single text STRING - wire it into any text node, a prompt input, or a show-text display.

Install

Ships in ComfyUI-Runware:

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

Restart (or ComfyUI Manager → search "Runware"). The model runs in OpenAI's cloud via Runware - no weights, no local LLM runtime. API key: Settings → Runware API key, RUNWARE_API_KEY, or runware auth login.

Gotchas

Pay-per-token, and thinking models are the ones where the bill creeps - xhigh thinking plus a long exchange adds up fast; the title bar shows cost after each run. There's no chat memory: each execution is stateless, so if you want multi-turn, you wire the output back into the Messages builder yourself. And if you only need fast, cheap text, the Mini/Nano variants of this node exist for exactly that - this one is for when you want the controls.