EvoLink Gemini Omni Flash Video Edit (Official)

The frustrating part of cloud video is that the generation step is only half the job. You render a clip, stare at it, and think "great, but swap the background to a mountain range." With EvoLink_GeminiOmniEdit you don't have to restart from scratch - you feed the finished video back in and tell Google's Gemini Omni Flash what to change. It's the same edit-in-place trick that makes image editors so nice, just for video.

This is Google's Omni Flash video-edit model, the budget-tier sibling of the Veo line, and it's genuinely handy for the "close enough, except…" iteration loop. Change the subject, restyle the whole thing, or alter a localized region - the model treats the input clip as the source and regenerates it per your instruction.

The inputs that matter

Only three are worth touching on a first run:

• prompt (required) - the edit instruction, in plain language. The tooltip's example says it all: "把背景换成雪山" ("change the background to a mountain range"). Chinese works, as does English.
• video_url - the video to edit. This is the field that makes the pack's pipeline story work: it accepts a URL directly, so you can wire the result_urls output of any upstream EvoLink generation node straight into it. Generate → edit → upscale, all in one graph.
• api_key - fill once, it's saved locally, leave blank after. timeout_seconds defaults to a generous 1800s, which is worth knowing: video edit jobs can legitimately take a while.

The outputs

Standard for every video node in this pack: video (the edited result, already downloaded to a local VIDEO object you connect to a native Save Video node), task_id (for the dashboard logs at evolink.ai/zh/dashboard/logs), status_info (human-readable summary with model, time, credits used), result_urls (raw links, one per line, 24-hour expiry), and response_json (the full task response for downstream parsing).

How it runs

Mechanically it's the same async dance as all the other nodes here. The pack submits the job to POST /v1/videos/generations with gemini-omni-flash-video-edit as the model, then polls GET /v1/tasks/{id} until the clip is done, updating a progress bar on the node. The video downloads to your disk when it completes - the raw links expire after 24 hours, which is why the node grabs it for you. Failed or content-blocked tasks don't charge you, so a refusal costs a few minutes, not money.

Installing and troubleshooting

No GPU, no model files, nothing heavy - the only dependency is requests, which ComfyUI already has.

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

Then fully restart ComfyUI (restart the console process, don't just refresh the browser), or install via ComfyUI Manager by searching EvoLink.

Watch for these: 401 means a bad key, 402 means out of balance (grab the EvoLink Credits node for a pre-check), 429 is rate limiting. If you're on a VPN, add api.evolink.ai and files-api.evolink.ai to your direct-connect rules - the pack's own client code calls this out, and it's a real failure mode. And if you get an error about comfy_api / VIDEO output, your ComfyUI is too old for video-output nodes; update ComfyUI itself first.