EvoLink Sora 2 (Official)

Sora has been the model people spend sentences about - the one from the demos, the one you could never run, the one that made a local Wan generation feel like getting away with something. EvoLink Sora 2 (Official) is the wrapper that drops OpenAI's Sora 2 into a ComfyUI canvas: text-to-video and image-to-video through a pay-per-use API, with nothing running on your machine.

How it works

It's the same architecture as the whole EvoLink pack - submit a job to /videos/generations, poll until the platform says it's done, download the clip before the 24-hour result links expire, and hand it back as a native VIDEO output that connects straight to ComfyUI's Save Video node. No GPU, no downloads, no ComfyUI bakeoff on your card.

One genuinely nice mechanism detail: Sora 2 requires the reference image to be exactly the pixel size of the chosen aspect ratio. Rather than making you get that right by hand, the node auto-resizes your image to 1280x720 (16:9) or 720x1280 (9:16) before it uploads. Load an image, pick a ratio, done - no "why does Sora keep rejecting my reference" forensics.

The inputs that matter

• prompt - required, up to ~5000 tokens, Chinese works. It's a camera/shot description, so write like a director's note.
• aspect_ratio - 16:9 landscape or 9:16 portrait, nothing else. Sora 2 doesn't do square.
• duration - 4, 8, or 12 seconds, and you're billed by the second, so this is your main cost lever.
• image - optional reference for image-to-video. Important: Sora's content moderation is strict, and the tooltip is blunt about it - no real-person photos as reference. Don't fight this; it refuses at the source and there's no local patch for a model you don't own.

Outputs

video is the one you wire into Save Video (the node needs a reasonably current ComfyUI for native VIDEO output - if you're on an old build you'll get an error telling you to upgrade). task_id is your receipt for the job log, status_info summarizes model/time/credits, result_urls has the raw links, response_json the full payload. And because everything EvoLink emits result_urls, you can chain the clip straight into the Topaz upscale node for a generate → upscale pipeline.

Install and gotchas

Part of the comfyui-evolink pack. Manager → search EvoLink → Install, then fully restart ComfyUI (close the console window - a browser refresh isn't enough). Or:

cd ComfyUI/custom_nodes
git clone https://github.com/deeplearning-goethe/comfyui-evolink

Only requests as a dependency; no model files, no VRAM. API key from evolink.ai/dashboard/keys, fill api_key once then leave it blank so it doesn't ride along in shared workflow JSON.

The default timeout_seconds is 1200 (20 minutes) - video takes a while, so don't panic if the queue shows progress for a long stretch. Failed or moderation-blocked tasks don't charge. 401 = bad key, 402 = no credit, 429 = rate limited. And if you rerun an unchanged workflow, ComfyUI caches it - Sora costs enough that the cache refund is a real feature, not a footnote.