WaveSpeedAI InfiniteTalk

Generate talking head videos from a portrait image and audio using WaveSpeed AI's InfiniteTalk model. Produces lip-synced video up to 10 minutes long with natural facial animations.

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 by @wavespeed/mcp (npx -y @wavespeed/mcp; tools search_models, get_model_schema, get_price, upload_file, run_model, get_prediction). It shares the CLI's stored login. Every example below maps one-to-one onto run_model with the same model id and input fields.

Quick Start

# Upload local image and audio files
OUTPUT_URL=$(wavespeed run wavespeed-ai/infinitetalk \
  -i image=@./portrait.png \
  -i audio=@./speech.mp3 \
  --json | jq -r '.outputs[0]')

Existing URLs work as-is:

OUTPUT_URL=$(wavespeed run wavespeed-ai/infinitetalk \
  -i image="https://example.com/portrait.jpg" \
  -i audio="https://example.com/speech.mp3" \
  --json | jq -r '.outputs[0]')

API Endpoint

Model ID: wavespeed-ai/infinitetalk

Animate a portrait image with lip-synced audio to produce a talking head video.

Parameters

Example

OUTPUT_URL=$(wavespeed run wavespeed-ai/infinitetalk \
  -i image=@./portrait.png \
  -i audio=@./speech.mp3 \
  -i resolution="720p" \
  -i seed=42 \
  --json | jq -r '.outputs[0]')

Using a Mask Image

When multiple people are in the image, use a mask to specify which face to animate:

OUTPUT_URL=$(wavespeed run wavespeed-ai/infinitetalk \
  -i image=@./group-photo.png \
  -i audio=@./speech.mp3 \
  -i mask_image=@./mask.png \
  -i resolution="720p" \
  --json | jq -r '.outputs[0]')

Important: The mask should only highlight the face region to animate. Using the full image as a mask will produce a fully black output.

With Prompt Guidance

OUTPUT_URL=$(wavespeed run wavespeed-ai/infinitetalk \
  -i image=@./group-photo.png \
  -i audio=@./speech.mp3 \
  -p "natural head movements, subtle expressions" \
  --json | jq -r '.outputs[0]')

Resolution and Pricing

Minimum charge is 5 seconds. Video length is determined by the audio duration (up to 10 minutes).

Tips

• Use a clear, front-facing portrait for best results
• Audio quality matters — use clean speech recordings with minimal background noise
• Keep prompts short and in English to avoid noisy or unexpected results
• For group photos, always provide a mask_image to target the correct face
• 480p is faster to generate; use 720p when higher quality is needed
• Processing time is approximately 10-30 seconds of wall time per 1 second of video

CLI tips

# Inspect the live input schema before running (fields, enums, defaults)
wavespeed run wavespeed-ai/infinitetalk -h

# Quote the price first
wavespeed price wavespeed-ai/infinitetalk -p "..." -i key=value

# Save outputs to disk instead of only printing URLs
wavespeed run wavespeed-ai/infinitetalk -p "..." --json --download "./out/{index}.{ext}"

# Local files: prefix the path with @ and the CLI uploads it and passes the hosted URL
wavespeed run wavespeed-ai/infinitetalk -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.

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.