Runway Text to Image

Runway is a name with history in this space - the company that published Stable Diffusion 1.5 back when it was still cozy with the open-weights movement, before it deleted its repos and went fully closed and commercial. Its Gen 4 model is the current hosted generation, and this node is the image half of it, exposed directly in the node graph. No Runway account juggling, no app, no web interface: prompt in, tensor out.

Gen 4 is a strong, cinematic-leaning model. Runway's whole brand is filmic output - it comes from the video world - and its stills carry that same look. Where it tends to get picked for is photographic, atmospheric generations rather than illustration. The tooltip on the ratio input hints at the same bias: this is a model built to produce frames.

Inputs that matter

• prompt - multiline, the whole job
• ratio - aspect ratio combo (1:1, 16:9, 9:16, and friends)
• optional reference_image - a single image to guide the generation

That's the entire schema - three inputs, one IMAGE output. No seed, no strength, no negative prompt, no style preset. This is a deliberately thin node: Runway keeps the knobs server-side.

How it works

It's a task-based API, not a one-shot call. Your prompt goes to /proxy/runway/text_to_image, Runway returns a task ID, and the node polls /tasks/{id} until the job finishes, then downloads the result into a tensor. That means runs take noticeably longer than the synchronous Recraft nodes - you're waiting on a queue, and the progress bar can sit for a bit. If you're batch-processing, budget the time.

If you provide a reference_image, it gets uploaded first and passed as a reference - that's the img2img-ish path, guiding composition or style from a source. There are constraints the node enforces: the reference must be under 8000px and its aspect ratio must fall between 1:2 and 2:1. Keep it roughly square-ish and you're fine.

Getting it

Ships with ComfyUI core - nothing to install, no model files, no Runway API key. Auth is handled through your Comfy account, and the call is billed from your Comfy credits.

Common issues

• Slow is normal. Task-based generation means queue time. Don't panic when it doesn't return in seconds; this is the nature of the integration.
• No knobs to fix a bad result. If the output misses, you rewrite the prompt or swap the ratio - there's no seed to hunt and no denoise to tune. Some people find that liberating; some find it frustrating.
• Reference image constraints. Oversized or very wide/tall references get rejected by the validation. Downscale or crop before you wire them in.
• Closed model, no weights. Everything about Gen 4 lives on Runway's servers. If the API changes, the node changes with it; you have zero local fallback beyond what the rest of your graph can do.

Reach for it when you want Runway's cinematic photographic look or its guided reference generation without leaving ComfyUI. It's a genuinely good hosted model with a minimal interface - just bring patience for the queue.