Runway Gen-4 Image

Runway's Gen-4 is the image side of the same company that owns a chunk of the AI-video conversation, and its image model has a specific reputation: it's extremely good at keeping a subject consistent across a batch of frames, and it takes reference images seriously. This node puts that model inside your graph, running on Runware's API - same pitch as the rest of the pack, no Runway subscription UI, no web tab, just a node with an IMAGE out.

The interesting thing about Gen-4 image generation is that it's built around reference conditioning rather than raw prompt luck. Where a typical text-to-image node hopes the prompt pins down the character, Gen-4's whole design assumes you'll hand it a reference and ask it to respect it. That makes this node a genuinely different tool in the Runware catalog than, say, a FLUX node.

Inputs that matter

positivePrompt is required; everything else is optional but Gen-4 only shows what it's for when you use the optional stuff:

• referenceImages - wire one or more IMAGE tensors in. This is the key Gen-4 capability: the model uses these to lock identity, style, or composition. Character consistency is the job this model is actually good at, so feed it a face or a character sheet and prompt from there.
• width / height - 1024×1024 defaults, plain pixel dimensions.
• numberResults - up to 20; each result uses a different seed. Batch variations are cheap here.
• providerSettings.runway.contentModeration.publicFigureThreshold - auto or low. This is Runway's own content policy leaking into the node: how strictly it filters for public figures. If you're generating a lookalike of a real person, this is the setting that decides whether you get a result or a polite block. low loosens it; don't assume that's a smart move.
• safety.checkContent - the pack-level content check, separate from Runway's own moderation above.

outputFormat (JPG/PNG/WEBP) and outputQuality are the usual file knobs. There's no seed input on this node - a fun quirk of the Gen-4 spec - so reproducibility rides on numberResults and the reference set.

How it works

Under the hood: the node fires a runway:4@1 image-inference request through the Runware SDK, your reference images get uploaded as part of the request, and the response URL is converted into a native IMAGE. Cost and content-check status print on the node's title bar after the run. Nothing runs locally, which is the entire appeal - Gen-4 is a hosted model you can't download anyway, so this node is the only way it's ever going to live inside a ComfyUI workflow.

Install and gotchas

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

Restart, add your Runware API key in Settings (or RUNWARE_API_KEY, or runware auth login). It's pay-per-use with a minimum top-up.

Where people get burned: they use it as a plain text-to-image node and wonder why it's not "better" than the local model they already run. Gen-4's edge is the reference path - if you're not wiring referenceImages, you're using the most expensive text-to-image node in the pack. Also, remember there are two moderation layers: Runway's publicFigureThreshold and the pack's safety.checkContent. If a run comes back empty, check both, and don't be surprised when "auto" threshold quietly kills a celebrity prompt.