slavasexton/comfyui-agent-kit/seedance

Seedance video prompting

ByteDance's Seedance family generates video and audio jointly and takes multimodal references (images, video clips, audio clips) that a prompt addresses by label. Getting the labels and the task type right matters more than adjectives: the same sentence can be read as "reference this" or "edit this" depending on one word.

Load `reference.md` for the full formulas, the failure table, worked examples and the ComfyUI parameters. Load `supplementary.md` only for non-official leads about 2.5; it never overrides reference.md and carries a conflict ledger. This file is the router and the rules you need before writing a single line.

Which model you are talking to

Seedance 2.5 reached ComfyUI on 2026-08-08 (core v0.31.0, PR 15395), and this file said the opposite until 2026-08-09. The old line, "Seedance 2.5 has no ComfyUI nodes", was correct on 2026-08-01 and went stale a week later. Kept as a correction rather than silently replaced.

How to build it now. 2.5 is a model option inside the SAME ByteDance2* nodes, not new nodes: ByteDance2TextToVideoNode, ByteDance2FirstLastFrameNode, ByteDance2ReferenceNode (category partner/video/ByteDance). Set the model DynamicCombo to Seedance 2.5 and the widgets underneath change. Minimal graph: the node's VIDEO output straight into SaveVideo.video.

What ComfyUI's 2.5 gives you, read from `comfy_api_nodes/nodes_bytedance.py` on master (2026-08-09), not from the announcement:

• resolution is 480p or 720p only. There is no 1080p and no 4K on 2.5 in ComfyUI; those live on

the Seedance 2.0 option of the same node.

• duration 4 to 30 s, ratio 16:9 / 4:3 / 1:1 / 3:4 / 9:16 / 21:9 / adaptive (the first-last-frame

node has no ratio widget), generate_audio on by default, output_format mp4 only even though the node's own model tooltip advertises "mp4/mov".

• ByteDance2ReferenceNode autogrows to 30 images, 10 videos, 10 audios, which confirms the

announcement's 30+10+10 in code. Its video_editing toggle makes the output inherit the source clip's duration and aspect, and the duration / ratio widgets are then ignored.

• Node-level prompting rule: put spoken lines in double quotes to steer the generated dialogue.
• The Long Video 30 to 180 s mode in the table above is a Dreamina / Jimeng feature and is not

reachable through these nodes, whose ceiling is 30 s. Nest-extend instead.

The three task types, and the word that switches between them

Every Seedance prompt is one of three tasks, and the model decides which from your phrasing:

The trap: for edit and extend, name the asset as `<Video_N>` directly. Writing "reference <Video_N>" makes the model treat it as a reference task and generate something new instead of editing what you gave it. (Confirmed, official BytePlus guide.)

Reference labels

Official ByteDance examples write `@Image 1`, `@Video 1`, `@Audio 1` with a space. Dreamina's consumer guide shows @Image1 without one. Both appear in the wild; the official form is spaced. Do not invent other label shapes.

Two ways to name a subject, and you must pick one and keep it:

• Bind inline: <Subject_N>@<Image_N>, e.g. Zhang San@Image 1, repeated every single time the

subject is mentioned.

• Define once, then reuse the label: "define the tall man in Video 1 as police officer", then

say "police officer" consistently forever after.

Define a subject with 2 to 3 stable static features (clothing, hairstyle, category), never with mood or action. Every mention must be explicit; an unlabelled mention is where subjects get swapped.

Four symbols that carry meaning

Officially documented, and nothing else does their job:

Non-Chinese, non-English dialogue must name its language: says in Japanese {こんにちは}.

Shot sequencing beats prose

Write a storyboard, not a paragraph. Label segments Shot 1, Shot 2, Shot 3 in the order events happen, and give each one: camera move or transition, subject action and expression, position or spatial change, audio.

Timing depends on the version, and this reversed in 2.5.

• 2.0: do not force timings. The official 2.0 guide says support for precise timing such as

"0 to 3 seconds" is unstable and constraining duration can break the generation.

• 2.5: timestamps are a headline feature, built because users asked for exactly this. Write time

slices as 0s-3s:, 3s-8s: and direct each one. See reference.md section 2.5-C.

One camera movement per shot. Asking for push, pull, pan and track at once destabilises the image.

Do not fill the reference budget

The single most counterintuitive official rule. 2.5 accepts 50 assets; the guide recommends 4 to 5: 1-2 character images (headshot plus full body) + 1 scene image + 1 camera-movement video + 1 audio clip. Too many assets and the model cannot rank features, producing style conflicts and blurred subjects.

Place the assets that need the most faithful reference earliest in the prompt.

When something comes out wrong

Full causes, the rest of the table and the ComfyUI-specific limits are in reference.md.

Sources and honesty

Normative content here comes from ByteDance primary sources: the official BytePlus ModelArk prompt guide for the Seedance 2.0 series, the official Seed blog announcing 2.5 (2026-07-31), Dreamina's official product guide, and ComfyUI's own node source. Third-party writing is kept out of the normative layer and lives in supplementary.md, subordinate and with its conflicts listed.

The detailed prompting rules are the official 2.0-series guide. Seedance 2.5 launched 2026-07-31 and has no public prompting guide of its own yet; it inherits the mechanics visibly, but treat 2.5-specific numbers as the only part confirmed for 2.5. reference.md marks every claim.