Muse Image

Most image generation nodes are dumb renderers: prompt in, pixels out, and whatever the model hallucinated about the Eiffel Tower is on you. Muse Image is not that. Read its settings and you'll see it's an agentic image model - it can search the web, pull visual references for real places and products, run code to build charts and structured layouts, and reason about the task before it renders. This node is Meta's current image model served through Runware's cloud, and the knobs it exposes tell you more than most model cards do.

Let the defaults show you the intended workflow. settings.webSearch is on - the generation is meant to be grounded in current facts, which is what turns "make a poster for a product launch" from a hallucination gamble into something that can check dates and names. settings.imageSearch is on too, pulling visual references for real places, products, logos, and styles - this is how the model renders something like the Apple logo correctly instead of inventing a plausible-but-wrong version. settings.shell lets it run code to build charts, tables, and structured layouts before rendering, which is the "make me an infographic" door. And settings.thinkingLevel (default high) controls how much internal reasoning happens before generation. If you want raw creative speed with no agentic preamble, dropping thinkingLevel and the search toggles is how you get closer to a plain text-to-image model.

The rest is familiar. positivePrompt is required, in sentences. referenceImages supplies an image reference, and there's an advancedFeatures.watermark.* group for stamped watermarks. size defaults to 2K (1:1) with a preset dropdown. settings.moderation defaults to auto - an explicit moderation control you won't find on most nodes. numberResults, the safety group, outputFormat (JPG default), and outputQuality close it out. Output is image.

Install

Muse Image is one node in the Runware pack, so setup is the pack setup: ComfyUI Manager → search Runware → install → restart, or clone https://github.com/Runware/ComfyUI-Runware into custom_nodes and pip install -r requirements.txt. Add a Runware API key from the dashboard - Settings → Runware API key, RUNWARE_API_KEY, or runware auth login. No models download; generation, searching, and code execution all happen on the hosted side.

What to know before you lean on it

Honesty corner: Muse Image is newer than this KB's coverage, so I can't hand you a community reputation for it the way I could for, say, Qwen - what's written here is read off the node's own controls. Two implications follow. First, the agentic features (web search, image search, code execution) are a genuinely different surface from a diffusion node: they cost more and take longer, so switch them off when a prompt is self-contained. Second, an agentic model is a data-in/out machine in a way a plain renderer isn't - your prompt can trigger real web and image lookups, and everything transits Runware's cloud and moderation. For proprietary material, that's a real consideration; for "make me a researched hero image," it's the feature.