Sora 2

OpenAI's Sora 2 is a closed model. You can't download the weights, you can't quantize it, and no amount of VRAM will ever run it locally. So if you want Sora footage inside a ComfyUI workflow, your only option is an API - and that's precisely the job this node does. It turns Sora 2 into a single node that takes a prompt and hands you a VIDEO tensor on the other end, no OpenAI SDK, no signing code, no upload step.

This is the appeal of the whole Runware pack in miniature. You're not generating locally; the node ships your request to Runware's cloud, Sora renders there, and the result comes back as a native ComfyUI video. Your GPU sits idle the whole time. For a model like Sora 2 that's not a compromise - it's the only way.

How it works

Under the hood, Runware_openai_sora_2 is a thin generated wrapper. The pack's base node class sends a videoInference request (model AIR openai:3@1) over REST via the runware-sdk, downloads the returned video file, and wraps it in ComfyUI's VIDEO type. Any IMAGE you wire into frameImages gets base64-encoded and sent as starting frames. You're watching an API call from inside the node graph, and that's the whole trick.

The inputs that matter

Only one thing is required, and it's the obvious one:

• positivePrompt - the text describing what to generate. This is Sora, so it can also express camera moves, cuts, and pacing ("slow push-in on a rain-soaked street" works far better than in most models).

The rest of the widget set is where you'll actually spend time:

• duration - a dropdown from 4 to 20 seconds, or (default). Total frames = duration × the model's fps, so longer is not just longer, it's costlier per run.
• frameImages - wire in an IMAGE to start generation from existing frames (image-to-video or a lead-in).
• videoId - paste the ID of a video you already generated to remix or extend it. This is how you do continuation chains.
• width / height - defaults to 1024×1024; the model likes the common ratios, so don't go exotic unless you have a reason.
• numberResults - 1 to 4, each a variation on a different seed.

The safety.* toggles and outputFormat (MP4/WEBM/MOV) with outputQuality are self-explanatory; leave them alone until you need them.

Installing it

You don't install Sora 2 - you install the pack that hosts it. In ComfyUI Manager, search Runware, install, restart. Or manually:

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

Then create a key at runware.ai/api-keys and paste it into ComfyUI Settings → Runware API key (or set RUNWARE_API_KEY). No model files download - Sora 2 lives entirely in Runware's cloud.

Where people get burned

• No key, no run. The node fails with "No Runware API key" if you forget this step.
• It costs money, every run. The node title bar shows a per-run cost after each generation - this is metered cloud inference, not local sampling. A 20-second Sora render is not free.
• Latency. A video takes real time in the cloud. Don't queue 20 variations and wonder why your graph looks hung; it's not hung, it's waiting on Sora.
• Old ComfyUI versions may hand you a file path instead of a VIDEO object - update ComfyUI if the output socket looks wrong.

If you want Sora 2 but a lower price point for drafts, the same pack's Runware_openai_sora_2_pro is the bigger, more expensive tier. For everything else, this is the node.