🫧 Seedance 2.0 Pro , Pro Pack on RunComfy

runcomfy.com · docs · Seedance 2.0 Pro model page

Original author: kalvinrv

intent

Seedance 2.0 Pro generates 4, 15 second cinematic short-form video clips at up to 720p with native in-pass lip-sync audio. use this skill when you need multi-modal video generation (text prompt + up to 9 image refs + up to 3 video refs + up to 3 audio refs), spokesperson dialogue with synchronized speech, or cinematic previs with explicit camera motion. route here if the user says "seedance", "seedance 2", "seedance v2", "seedance pro", "seedance 2.0", "bytedance seedance", or explicitly asks for seedance video generation.

inputs

required

• prompt (string): seedance accepts either chinese (max 500 chars) or english (max 1000 words). include camera direction ("medium close-up", "slow push-in", "handheld follow", "locked tripod"), lighting, mood, and audio tone. text controls narrative and motion; stable identity goes in image_url instead.

optional

• image_url (array of urls): 0, 9 image references (jpeg/png/webp/bmp/tiff/gif). use for stable identity (face, costume, logo, brand mark). seedance preserves these across generations.
• video_url (array of urls): 0, 3 video references (mp4/mov), each 2, 15 seconds. use for scene context, environment, or motion reference.
• audio_url (array of urls): 0, 3 audio references (wav/mp3), each 2, 15 seconds and under 15mb. use for voice tone, accent, or ambient reference. seedance treats these as guides, not strict drivers.
• aspect_ratio (enum): adaptive (default, matches input refs), 16:9, 9:16, 4:3, 3:4, 1:1, 21:9.
• duration (integer): 4, 15 seconds (default 5). seedance hard cap is 15s per call.
• resolution (enum): 480p or 720p (default 720p).
• generate_audio (boolean): default true. when true, seedance synthesizes lip-synced speech / sfx / music in the same pass as video. set false to get video-only output.
• seed (integer): optional. pass a fixed seed for reproducible output. same prompt + same seed = deterministic generation.

external connections

• RunComfy CLI: npm i -g @runcomfy/cli. required locally or in ci.
• RunComfy account & token: run runcomfy login once (opens browser device-code flow). stores api token to ~/.config/runcomfy/token.json (mode 0600).
• RUNCOMFY_TOKEN env var (ci/containers): set this instead of runcomfy login. token has no expiry but reject if invalid.
• Network: cli connects to model-api.runcomfy.net, polls every 2 seconds, downloads results from *.runcomfy.net or *.runcomfy.com.

procedure

1. validate runcomfy cli and auth

   • input: system $PATH, ~/.config/runcomfy/token.json or $RUNCOMFY_TOKEN
   • action: check runcomfy --version succeeds. if not, user must install via npm i -g @runcomfy/cli.
   • action: check auth exists. if missing, user runs runcomfy login (interactive) or sets $RUNCOMFY_TOKEN (ci).
   • output: confirmed runcomfy cli is callable and authenticated.

2. build seedance input json

   • input: user's prompt, optional image/video/audio urls, optional aspect_ratio/duration/resolution/seed.
   • action: construct json object with prompt (required) and optional fields.
   • action: validate prompt length: if chinese, reject if > 500 chars; if english, reject if > 1000 words (rough: word_count = prompt.split().length).
   • action: validate image_url array length <= 9. validate video_url array length <= 3 and each url valid. validate audio_url array length <= 3 and each url valid.
   • action: validate duration is int 4, 15. validate resolution is "480p" or "720p". validate aspect_ratio is one of the 7 enum values. validate generate_audio is bool.
   • output: json string ready for --input flag, or error if validation fails.

3. invoke runcomfy cli

   • input: validated input json from step 2, user-specified absolute output directory path.
   • action: run runcomfy run bytedance/seedance-v2/pro --input '<json>' --output-dir <path>.
   • action: cli immediately prints request id and begins polling every 2 seconds.
   • output: cli output shows polling progress or error message.

4. handle polling and timeout

   • input: cli process running, polling the runcomfy backend.
   • action: monitor cli output and exit code. seedance generation typically finishes within 30, 90 seconds depending on duration and references.
   • action: if user presses Ctrl-C, cli cancels the remote seedance request and exits cleanly.
   • action: if cli does not return within 5 minutes, treat as timeout (network/backend issue). user can retry.
   • output: either generation succeeds and downloads video file(s), or cli exits with an error code.

5. download and verify output

   • input: cli process completed (exit code 0 or nonzero), output-dir specified in step 3.
   • action: cli automatically downloads any .runcomfy.net or .runcomfy.com url into the output-dir.
   • action: check output-dir for video files (typically named with timestamp and extension .mp4 or .mov).
   • action: if exit code is 0, video generation succeeded.
   • output: mp4/mov video file in output-dir, ready for playback or further processing.

decision points

if the user has no runcomfy cli installed:

• user runs npm i -g @runcomfy/cli and retries.

if the user has no runcomfy token (interactive environment):

• user runs runcomfy login, which opens a browser and stores token to ~/.config/runcomfy/token.json.

if the user is in ci/container and has no token:

• user sets RUNCOMFY_TOKEN=<token> env var and retries.

if the prompt is too long (chinese > 500 chars or english > 1000 words):

• reject the input and ask user to shorten the prompt.

if image_url array > 9, video_url array > 3, or audio_url array > 3:

• reject and ask user to reduce the number of references. seedance has hard caps.

if duration is outside 4, 15 seconds:

• reject and ask user to pick a duration in range, or split longer narratives into multiple seedance calls.

if cli returns exit code 77 (not signed in or token rejected):

• token is invalid or expired (or missing). user must re-run runcomfy login or update $RUNCOMFY_TOKEN.

if cli returns exit code 65 (bad input json):

• input schema mismatch. review the json for typos in field names or invalid enum values. check prompt length.

if cli returns exit code 69 (upstream 5xx):

• runcomfy backend is down or experiencing issues. retry after 30, 60 seconds.

if cli returns exit code 75 (timeout or 429 rate limit):

• either the backend is slow or you hit a rate limit. retry after 60 seconds. if repeated, back off exponentially.

if cli returns exit code 64 (bad cli args):

• user passed invalid flags or malformed json. check the command syntax.

if the output-dir does not exist:

• cli will create it. if creation fails (permissions), user must ensure write access to the target directory.

if generate_audio is true but you want video-only output:

• set generate_audio: false in the input json.

if you need deterministic output (testing, variant comparison):

• pass a fixed seed int in the input json. same seed + same prompt = same video.

if reference audio is longer than 15 seconds or > 15mb:

• seedance rejects it. user must trim or re-encode the audio.

if reference video is shorter than 2 seconds or longer than 15 seconds:

• seedance rejects it. user must re-encode to the 2, 15s window.

if you want to describe a person's face or appearance in detail:

• don't put it in the prompt. instead, pass a photo as image_url. seedance preserves image identity better than text description.

if you're mixing conflicting aesthetic refs (watercolor + photoreal):

• simplify. seedance performs best with coherent visual language. remove the contradictory ref or reword the prompt.

if you need video longer than 15 seconds:

• seedance caps at 15s per call. split your narrative into multiple seedance calls and stitch the outputs in post.

output contract

• success: cli exits with code 0. an mp4 or mov file appears in the output-dir specified via --output-dir. file size is typically 5, 50mb depending on duration and resolution. file is ready for playback in any standard video player.
• file location: <output-dir>/<timestamp-or-id>.<ext> (cli decides the exact filename).
• file format: h.264 video codec, aac audio (when generate_audio: true), container format mp4 or mov.
• duration: matches the duration param passed (4, 15 seconds).
• resolution: matches the resolution param: 480p or 720p.
• audio: synchronized lip-sync (when generate_audio: true), or silent (when false).
• frame rate: typically 24fps.
• total file size cap: 2gb (enforced by runcomfy; individual seedance outputs well under this).

outcome signal

• user hears/sees the output: open the mp4/mov file in a video player. the video shows the seedance generation result. if generate_audio: true, audio plays back in sync with lip movement.
• cli feedback: final line of cli output says "success" or "complete" with a local file path.
• shell exit code: echo $? returns 0.
• file timestamp: output-dir contains a newly-created video file with today's timestamp.
• no errors in cli output: cli printed no warnings or errors; polling completed without timeout or 5xx backend errors.

---

Appendix: prompting best practices

image vs text division , the core rule. stable identity (face, costume, logo) goes in image_url. seedance preserves it exactly. evolving narrative (action, lighting, mood, camera move) goes in the prompt. seedance generates it. trying to describe a face in text wastes tokens and drifts.

camera and motion in plain language. seedance 2.0 pro understands "medium close-up", "slow push-in", "handheld follow", "locked-off wide" as real directives. combine: "medium close-up. slow push-in over 3 seconds. handheld, slight breathing motion." seedance executes the grammar.

audio direction with generate_audio true. tell seedance the tone: "warm friendly conversational", "calm instructional", "crisp newsroom delivery". for ambient: "gentle cafe chatter, distant traffic, no foreground music". seedance synthesizes audio matching the directive.

reference media specs. reference videos must be 2, 15s; reference audio must be under 15mb and 2, 15s. out-of-range files reject. match aspect ratio of refs to the seedance output to avoid crops.

anti-patterns to avoid:

• mixing radically different aesthetic refs (watercolor + photoreal). confuses seedance.
• conflicting style cues in the prompt. simplify by removing contradictions.
• trying to describe stable identity verbally. use image_url instead.
• asking seedance for > 15s clips. 422 error. segment into multiple calls.

Sample verified prompts

default example:

Golden hour on a quiet cafe terrace: a barista wipes the counter,
then looks up and explains today's special in a friendly tone, natural
lip-sync. Medium close-up, slow push-in; warm side light, soft bokeh
through glass, gentle cafe ambience and subtle film grain.

multi-modal lip-sync (text + image):

Same person as image 1 in a softly-lit recording booth, leaning into
the mic, says: "We just shipped the biggest update of the year."
Calm conversational tone. Medium close-up, locked tripod, shallow DOF,
warm key light from camera-left.

FAQ

max clip duration? 4, 15 seconds per seedance call. for longer narratives, segment and stitch.

aspect ratios? seven: adaptive, 16:9, 9:16, 4:3, 3:4, 1:1, 21:9. default adaptive (matches input refs).

does seedance do lip-sync? yes. with generate_audio: true (default), seedance produces lip-synced speech in-pass.

can seedance take existing audio as input? yes, pass as audio_url. seedance treats it as a reference (voice tone, mood) rather than a strict lip-sync driver.

languages? chinese (max 500 chars) or english (max 1000 words).

resolution ceiling? 720p on this endpoint.

reproducible output? pass a fixed seed int.

Limitations

• duration cap 15s per call.
• resolution cap 720p.
• reference media specs: video/audio 2, 15s; audio under 15mb.
• lip-sync depends on prompt clarity. not guaranteed perfect under all conditions.
• no @-syntax for character binding. relies on image refs + prompt alignment.

Exit codes

code	meaning
0	seedance generation succeeded
64	bad cli args
65	bad input json or schema mismatch
69	upstream backend 5xx error
75	retryable: timeout or 429 rate limit
77	not signed in or token invalid/rejected

full reference: docs.runcomfy.com/cli/troubleshooting.

How it works under the hood

the skill invokes runcomfy run bytedance/seedance-v2/pro with a json body matching the seedance schema. the cli posts to https://model-api.runcomfy.net/v1/models/bytedance/seedance-v2/pro, polls every 2 seconds, fetches the result, and downloads any .runcomfy.net or .runcomfy.com url into the output-dir. ctrl-c cancels the remote request cleanly before exit.

Security and privacy

• token storage: runcomfy login writes the api token to ~/.config/runcomfy/token.json with mode 0600. in ci, set $RUNCOMFY_TOKEN env var.
• input boundary: seedance prompt is passed as json via --input. no shell injection risk.
• third-party content: image/video/audio urls are fetched by the runcomfy server. treat external urls as untrusted. image-based prompt injection is a known risk for any image/video model.
• outbound endpoints: only model-api.runcomfy.net and *.runcomfy.net or *.runcomfy.com.
• generated-file size cap: 2gb per seedance call.