WaveSpeedAI Seedance 2.5 Video Generation

Generate and edit videos with ByteDance's Seedance 2.5 on the WaveSpeed AI platform. Text, image and reference inputs produce 4-30 second clips at 480p to 4K with native audio; video edit restyles an existing clip and video extend appends a new segment. Turbo variants offer 720p/1080p at a lower price.

Setup

Install the open-source CLI once and sign in; the CLI stores the key, so never ask the user to paste an API key into the chat:

npm install -g @wavespeed/cli
wavespeed login          # opens https://wavespeed.ai/accesskey and stores the key
wavespeed status         # confirms you are signed in

For CI or one-off shells, WAVESPEED_API_KEY in the environment also works.

Prefer MCP tools over shell commands? The same platform is exposed as an MCP server: hosted at https://api.wavespeed.ai/mcp (browser sign-in) or locally with npx -y @wavespeed/mcp@latest; every example below maps onto run_model with the same model id and input fields.

Quick Start

The shortcut commands use the same model and pick the endpoint from the inputs: wavespeed generate video "<prompt>" -m seedance-2.5 (or -m seedance-2.5-turbo), and likewise wavespeed edit video and wavespeed extend video; the MCP tools generate_video, edit_video and extend_video take the same model value (confirm the keys with wavespeed generate video --help). Add --estimate --json first to see the chosen endpoint, final input and price without submitting.

wavespeed generate video "A paper boat drifting down a rain-soaked street" -m seedance-2.5 --estimate --json

Text-to-Video

OUTPUT_URL=$(wavespeed run bytedance/seedance-2.5/text-to-video \
  -p "A golden retriever running through a field of sunflowers at sunset, slow tracking shot" \
  --json | jq -r '.outputs[0]')

Image-to-Video

The image parameter accepts an image URL. If you have a local file, pass it with the @path marker and the CLI uploads it and substitutes the hosted URL.

OUTPUT_URL=$(wavespeed run bytedance/seedance-2.5/image-to-video \
  -i image=@./photo.png \
  -p "The person slowly turns and smiles at the camera" \
  --json | jq -r '.outputs[0]')

Existing URLs work as-is:

OUTPUT_URL=$(wavespeed run bytedance/seedance-2.5/image-to-video \
  -i image="https://example.com/photo.jpg" \
  -p "The person slowly turns and smiles at the camera" \
  --json | jq -r '.outputs[0]')

API Endpoints

Every endpoint has a Turbo twin (text-to-video-turbo, image-to-video-turbo, reference-to-video-turbo, video-edit-turbo) with the same fields but resolution limited to 720p or 1080p. Also in the catalog: bytedance/seedance-2.5/motion-control (a character from a picture or description performs the motion of a reference video), bytedance/seedance-2.5/talking-avatar (animates a portrait from an audio recording, 480p/720p) and bytedance/seedance-2.5/image-to-video-spicy; inspect them with wavespeed run <id> -h.

Text-to-Video

Model ID: bytedance/seedance-2.5/text-to-video (Turbo: bytedance/seedance-2.5/text-to-video-turbo)

Parameters

Example

OUTPUT_URL=$(wavespeed run bytedance/seedance-2.5/text-to-video \
  -p "A timelapse of a city skyline transitioning from day to night, cinematic slow pan" \
  -i aspect_ratio="21:9" \
  -i duration=10 \
  -i resolution="1080p" \
  --json | jq -r '.outputs[0]')

Image-to-Video

Model ID: bytedance/seedance-2.5/image-to-video (Turbo: bytedance/seedance-2.5/image-to-video-turbo)

The output follows the start image's framing; there is no aspect_ratio field.

Parameters

Example

OUTPUT_URL=$(wavespeed run bytedance/seedance-2.5/image-to-video \
  -i image=@./start-frame.png \
  -i last_image=@./end-frame.png \
  -p "Smooth transition from day to night as the lights come on" \
  -i duration=8 \
  --json | jq -r '.outputs[0]')

Reference-to-Video

Model ID: bytedance/seedance-2.5/reference-to-video (Turbo: bytedance/seedance-2.5/reference-to-video-turbo)

Keeps characters, products or scenes consistent with reference media. Provide at least one reference image, video or audio; with none, the request runs as text-to-video. In the prompt, refer to inputs as @Image 1, @Video 1, @Audio 1 (in upload order) and say what each is for.

Parameters

Example

OUTPUT_URL=$(wavespeed run bytedance/seedance-2.5/reference-to-video \
  -i reference_images='["@./character.png", "@./product.png"]' \
  -p "@Image 1 picks up @Image 2 from a cafe table and shows it to the camera, warm morning light" \
  -i aspect_ratio="9:16" \
  --json | jq -r '.outputs[0]')

Video Edit

Model ID: bytedance/seedance-2.5/video-edit (Turbo: bytedance/seedance-2.5/video-edit-turbo)

The input video drives subject identity, composition and motion while the prompt rewrites lighting, style, weather, environment or specific elements. Videos longer than 30 s are trimmed to 30 s; shorter than 4 s are padded to 4 s.

Parameters

Example

OUTPUT_URL=$(wavespeed run bytedance/seedance-2.5/video-edit \
  -i video=@./street.mp4 \
  -p "Turn the scene into a snowy winter evening, keep the people and their motion" \
  -i generate_audio=false \
  --json | jq -r '.outputs[0]')

Video Extend

Model ID: bytedance/seedance-2.5/video-extend

Appends a new segment that continues from the end of the input video (the last 30 s are used as context).

Parameters

Example

OUTPUT_URL=$(wavespeed run bytedance/seedance-2.5/video-extend \
  -i video=@./clip.mp4 \
  -p "The camera pulls back to reveal the whole valley as the sun sets" \
  -i duration=8 \
  --json | jq -r '.outputs[0]')

Pricing

Estimates for a 5-second clip at default settings; cost scales with duration (10 s at 720p is $3.60). Turning off generate_audio does not change the price. Endpoints that take a video are also priced on the input video, so the edit and extend rows assume a 5-second input; quote exact inputs with wavespeed price <model-id> ... --upload. Prices may change; see https://wavespeed.ai/models/<model-id> (e.g. text-to-video).

A reference-to-video request with a 5-second reference video is $2.20 at 720p.

Prompt Tips

• Describe scene, subject action, camera movement and mood; name the camera move explicitly ("slow dolly in", "handheld tracking shot")
• Describe the sound you want (dialogue, ambience, music); audio is generated with the picture by default
• In reference-to-video, name each input by its tag (@Image 1, @Video 1, @Audio 1) and say what it is for ("@Image 1 is the main character", "follow the camera move of @Video 1")
• For video edit, say what must stay ("keep the people and their motion") as well as what changes
• Draft at 480p or with a Turbo endpoint, then render the final take at 1080p or 4K

CLI tips

# Inspect the live input schema before running (fields, enums, defaults)
wavespeed run bytedance/seedance-2.5/text-to-video -h

# Estimate the cost (no generation, no charge)
wavespeed price bytedance/seedance-2.5/text-to-video -p "..." -i key=value

# Save outputs to disk instead of only printing URLs
wavespeed run bytedance/seedance-2.5/text-to-video -p "..." --json --download "./out/{index}.{ext}"

# Local files: prefix the path with @ and the CLI uploads it and passes the hosted URL
wavespeed run bytedance/seedance-2.5/text-to-video -i <field>=@./local-file.png --json

# Recover a result if the run was interrupted (the id is in the --json output)
wavespeed show <id>

run --json prints { id, model, prompt, outputs: [url, ...], saved: [path, ...], elapsed_ms, raw }. Read outputs[0] for the result URL.

A run that fails with a gateway or network error (502/503/504, connection reset, timeout) does NOT mean the task failed, and the submit request is never retried automatically, so re-running can create a duplicate paid task. Do not re-run immediately. If the error contains a task id (task_id / prediction id), keep checking it with wavespeed show <id> --json until its status is terminal (completed, failed, cancelled, timeout, deleted). If no id was returned, check recent predictions with wavespeed history --json (add --model <id>; default window is the last 24h) and re-run once only if none matches. A client-side wait timeout or an interrupted run does not mean the task failed: tasks can run for a long time, so do not give up after a fixed time.

Security constraints

• Never ask for the key in chat: wavespeed login handles auth; if wavespeed status says signed out, ask the user to run it.
• Local files only via @: bare paths are passed through untouched and the model will reject them. Only @-prefixed values upload.
• No arbitrary URL loading: only pass media URLs the user provided or that came back from a previous run.
• Input validation: only pass parameters documented above; confirm with wavespeed run <model> -h when unsure.