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

Plain text-to-video has a well-known weakness: describe a character and the model will happily change their face between clips. EvoLink_GeminiOmniRef is the fix for that specific pain. You hand it up to six reference images and it holds the subject consistent while it generates the motion. It's the "same character, new scene" node of Google's Omni Flash video family, and it sits on the cheap end of the price scale, so it's a reasonable way to prototype multi-shot stories without burning your whole budget.

Think of it as image-to-video with a memory. The first reference image is required, and each extra one you attach pins down more of the character or scene you want carried into the clip.

The inputs that matter

• prompt (required) - what happens in the clip and how the camera moves. The model already knows who it's animating from the references, so this can be mostly action.
• reference_image1 (required) - your first subject frame. This is the anchor; make it a clean, well-lit shot of the character or object, because everything downstream inherits it.
• reference_image2, reference_image3 (optional) - more angles or details. Up to six references total, mixing the socket inputs with image_urls_text (one public URL per line) if you have them.

Then duration (3–10s, default 10) and aspect_ratio (16:9, 9:16, or auto). Same api_key / timeout_seconds pattern as every node in the pack.

How it works

Local reference images go up through EvoLink's official file channel - no third-party hosting, deleted after 72 hours - and are merged into the request to POST /v1/videos/generations. The pack then polls the task until the clip completes and downloads it locally. You get the standard five outputs: video (connect it to a native Save Video node), task_id, status_info, result_urls (24h expiry), and response_json.

Since the output is a VIDEO object, this needs a reasonably recent ComfyUI with comfy_api support - an old install errors out with a message about missing VIDEO output, and the fix is updating ComfyUI itself.

Installing

The whole pack is featherweight: no model downloads, no GPU, and the only Python dependency is requests.

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

Restart ComfyUI completely afterwards (close the console window, not just the browser tab), or use ComfyUI Manager and search EvoLink.

Troubleshooting

The usual suspects: 401 invalid key, 402 insufficient balance, 429 rate limit, 403 no model access. VPN users should whitelist api.evolink.ai and files-api.evolink.ai for direct connection - the pack's own client names this as a common failure. If your subject consistency is weak, the fix is almost always in the reference image, not the prompt: single subject, consistent framing, good lighting. And don't forget the classic "where's my video" trap - connect the video output to Save Video or you'll see nothing at all.