EvoLink Gemini Omni Flash Text-to-Video (Official)

If you've never generated video because local models need a serious GPU, this is the node that gets you from zero to moving image in one step. EvoLink_GeminiOmniT2V is Google's Gemini Omni Flash text-to-video, exposed as a ComfyUI node: type a description, run the queue, and a few minutes later you have a 3–10 second clip sitting in your output folder. No GPU, no weights, no codec circus - it all happens on Google's infrastructure and comes back as a plain VIDEO you can save with the native node.

This is the plainest member of the Omni Flash family, and the one to start with if you're new to the whole cloud-video thing. It's cheap, it's fast, and its job is narrow: prompt in, clip out.

The inputs that matter

Honestly, just two:

• prompt (required) - describe the shot and the motion. The tooltip suggests camera/frame language ("镜头/画面描述"), so think like a storyboard: subject, action, camera movement, mood. Chinese works.
• aspect_ratio - 16:9, 9:16, or auto (default 16:9). For vertical shorts you'll want 9:16; auto lets the model pick.

Plus duration (3–10s, default 10), and the pack-wide api_key (fill once, saved locally, leave blank after) and timeout_seconds (default 1200 - video jobs can be slow; don't panic if the progress bar crawls).

How it works

The node posts your prompt to POST /v1/videos/generations, then polls the task endpoint until the clip finishes, updating a progress bar on the node as it goes (pending → processing → completed). When it's done the video is downloaded to your machine as a VIDEO object - that's why you must connect the video output to a native Save Video node to actually keep it. Alongside it you get task_id (to look up the job in the EvoLink task logs), status_info (model, runtime, credits burned), result_urls (the raw links, one per line, 24-hour expiry), and response_json for any downstream parsing.

Failed or content-blocked jobs don't charge you, and re-running unchanged parameters reuses ComfyUI's cache, so iterating on a prompt is cheap - which is exactly how you should use a budget model like this: throw ideas at it, keep only the keepers.

Installing

No model files, no GPU requirement, only requests:

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

then fully restart ComfyUI. Or search EvoLink in ComfyUI Manager. If the node doesn't show up, check for a nested folder (the pack's __init__.py must sit directly in custom_nodes/comfyui-evolink/) and confirm you restarted the process, not just the tab.

Troubleshooting

401 bad key, 402 empty balance (use the EvoLink Credits node to check first), 403 no access to this model, 429 rate limited. VPN users: add api.evolink.ai and files-api.evolink.ai to your direct-connect list. And if you see an error about comfy_api or VIDEO output, your ComfyUI is too old - update ComfyUI, this pack needs the newer video-output plumbing.