EvoLink Seedance 2.5 Reference-to-Video (Official)

Most video nodes take one image or nothing. Seedance 2.5 Reference-to-Video is the one that takes everything: up to 30 reference images, up to 10 reference video clips, up to 10 audio references, all mixed in one render, and it builds a video that honors the whole pile. EvoLink Seedance 2.5 Reference-to-Video (Official) is ByteDance's multi-modal mode as a cloud node - the closest thing a closed API gives you to "character consistency" and "style transfer" in one shot.

What it is

• prompt (required) - the shot description; you can reference the materials you feed it. Chinese welcome.
• image1–image4 - image reference slots (up to 30 when combined with the URL box). Feed a character, a style sample, a scene.
• video_urls - reference clips, one URL per line, up to 10 (each 2–30 seconds). This is where the pack's chaining magic comes in: you can paste another EvoLink node's result_urls output straight in.
• audio_urls - reference audio, up to 10 entries, to steer mood or sync.
• duration - 4–30 seconds, default 5, billed by duration.
• quality - 480p / 720p / 1080p.
• aspect_ratio - adaptive default, or the usual menu.
• generate_audio - on by default, free.

The only real requirement: at least one piece of reference material, image or video or audio. You don't strictly need all three - that's the point of the "at least give it something" design. Outputs: video, task_id, status_info, result_urls (24-hour expiry), response_json. Save Video node on the video output or you won't see the result.

Install and key

Standard EvoLink install. ComfyUI Manager → search EvoLink → install → full restart. Manual:

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

Only requests as a dependency; no models, no GPU. Key from evolink.ai/dashboard/keys (sk-), pasted once - saved locally, blank after, never in a shared workflow JSON.

The workflow that makes it special

This node exists to be chained. Generate a first pass with Seedance T2V, feed its result_urls into this node's video_urls, add reference images for the subject, and you get a coherent follow-up instead of a coin flip. Reference images uploaded from disk ride EvoLink's channel (auto-deleted after 72 hours) - no third-party image host needed. Reference URLs from an upstream node are the key trick: they're plain text, so they flow anywhere.

The catch is the usual multi-modal one: more references means more ways for the model to get confused, and Seedance's meter doesn't care whether you're testing. Keep the reference pile focused - a character image and one style clip will outperform a twenty-item grab bag. And the standing cloud deal applies: per-render cost, your materials leave the machine, failed and moderation-blocked tasks don't charge.