Seedance 2.0 Fast (Replicate)

Seedance 2.0 Fast is ByteDance's video generator on Replicate, and this node is the most stacked video wrapper in the pack: first-frame, last-frame, up to 9 reference images, up to 3 reference videos for motion transfer, up to 3 reference audio clips for lip-sync and audio-driven generation, plus native audio with dialogue. That's a lot of doors into one model.

The name "Fast" matters. Seedance 2.0 Fast is the speed tier of a closed model that was already controversial when it arrived in ComfyUI-land - the community reaction to Seedance-as-API was immediate and split: people wanted it in their workflow, and the same people winced at the per-video price when it landed. ByteDance runs the two-track posture: flagships closed and API-only, everything underneath open. Seedance is firmly the former. You're renting it per clip, and video clips are where these bills get real.

How it works

Pack standard engine around bytedance/seedance-2.0-fast. Images go up as PNG, audio as WAV, and the result downloads as an MP4 saved to output/video/ with video_path returned as a STRING. One mechanism note: reference_videos is a STRING input - you supply URLs, not uploads. Replicate's workers need fetchable files, so a local path won't work.

The inputs that matter

• prompt (required) - describe the video. You can reference your references in the prompt as [Image1], [Video1], [Audio1], etc., which is how you tell Seedance which character/style/voice to use where.
• image - first frame for i2v. Can't be combined with reference images.
• last_frame_image - last frame, only works if a first frame is provided, and also can't be combined with references.
• reference_images - up to 9, for character consistency, style guidance, scene composition. This is the character-consistency workhorse input.
• reference_videos - up to 3 URL strings (total ≤ 15s) for motion transfer and style reference.
• reference_audios - up to 3 AUDIO inputs (total ≤ 15s) for audio-driven generation and lip-sync; needs at least one reference image or video.
• duration - 1–15s, default 5. The trick: set -1 for "intelligent" duration and let the model pick the best length.
• resolution - 480p or 720p (default). No 1080p on the Fast tier.
• aspect_ratio - 7 ratios plus adaptive (let the model choose from your inputs). Default 16:9.
• generate_audio - on by default, and it's a doozy: synchronized dialogue (put spoken words in double quotes in the prompt), sound effects, and background music.
• seed - set for reproducibility.

Output: video_path, a STRING.

Install

cd ComfyUI/custom_nodes
git clone https://github.com/Pharma-Lobby/ComfyUI-Replicate-Select

Or ComfyUI Manager → Install Custom Nodes → "Replicate Select", restart. Dependency: replicate>=1.0.7. Token via environment variable:

export REPLICATE_API_TOKEN=r8_...

Where people get burned

• References are mutually exclusive. First/last-frame setup and reference images/videos are two different modes; mix them and the model errors.
• reference_videos takes URLs, not files. A common first-run mistake. Host the clips somewhere fetchable first.
• Dialogue is a quoting trick. Spoken lines need double quotes inside the prompt - "close your eyes" in the text tells it to generate that speech. Plain prose becomes narration/ambient audio instead.
• The bill is the feature. The community's honest complaint about Seedance-as-API was cost, and it stands. Draft at 480p, short durations, and treat "intelligent" duration (-1) with care - it tends toward the longer, pricier end.