MyteamOne · Seedance 2.0 (Multi-Reference)

The newest and most interesting node in the pack. Where I2V gives you one first frame, Multi-Reference hands Seedance 2.0 a whole reference kit at once - up to five images, a reference video, and a reference audio track - and asks it to weave them into one scene. Think "this character from image 1, in that setting from image 2, following the camera move from this clip, with that mood of music." It's the closest thing the pack has to production-grade control from inside ComfyUI, which is why the README's intended use is drama production and why the default prompt is in Chinese (a character conversation in a warm interior).

How it works

Wire up to five ComfyUI IMAGEs into image_1 through image_5. Each gets base64-encoded as a JPEG and dropped into the media[] payload. Optional video_url and audio_url add motion and sound references - but those two must be hosted HTTP/HTTPS URLs, because base64 video/audio isn't supported by this API. Images are the only thing that travels as a data URI. Give it nothing at all and it fails fast, pointing you at the T2V node.

Here's the quirk that trips everyone: you describe the references in natural language, not by ordinal. Seedance 2.0's multi-ref mode doesn't understand "图1/图2"-style labels. The README's advice is to write something like "参考第一张图的角色，在第二张图的场景里…" - "the character from the first image, in the setting from the second" - and let the model match them. The images are positional inputs, but the prompt is where you tell the model what role each one plays.

The inputs that matter

Required: client, prompt, model (defaults to standard Seedance 2.0 here, not Fast - the multi-ref path leans on the bigger model), resolution, ratio (defaults to 9:16, so portrait is the default composition), duration, generate_audio, watermark, seed. Optional: the five images, video_url, audio_url. Output is one video, plus the saved myteamone_seedance_<taskid>.mp4 in ComfyUI/output/.

Install

Shared with the whole pack - clone once, no models to download:

cd ComfyUI/custom_nodes
git clone https://github.com/agione-pro/ComfyUI-MyteamOne
cd ComfyUI-MyteamOne
pip install -r requirements.txt

Restart; it lives under Add Node → MyteamOne/Seedance2.0 (the only node with its own category). Or install via ComfyUI Manager, searching "MyteamOne".

Where it falls over

• Dirty references reproduce their dirt. Feed it storyboard frames with arrows, callouts or labels and the model faithfully renders those annotations as on-screen graphics in the output. Bare character/scene art only.
• Real faces are blocked. InputVideoSensitiveContentDetected.PrivacyInformation - ByteDance's policy kills any video containing real, identifiable people via third-party APIs. AI-generated characters end-to-end, per the README.
• Fast + 1080p fails fast, same as everywhere in this pack. 720p with Fast, or standard Seedance for 1080p.
• Video and audio references need hosted URLs. A local file won't upload. Stick the clip on Aliyun OSS, Tencent COS, or Qiniu and paste the direct link.
• It's the priciest node here. More inputs, bigger model - multi-reference is the most expensive thing in the pack per call. Test prompts at 480p before spending on the full kit.