The Inference Gateway CLI can render short video clips through the TextToVideo agent tool, in two modes:

• Prompt render - a text prompt in, an MP4 out. Optionally pass a portrait as the first frame, or an avatar from the library as reference images so the same person shows up in the clip.
• Avatar render (lip-sync) - a portrait plus a .wav or .mp3 clip in, an MP4 of that face speaking the audio out. Pair it with Text-to-Speech to go from a script to a talking-head clip without leaving the chat.

Rendering goes through the gateway's Videos API (POST /v1/videos, then polling GET /v1/videos/{id} until the job finishes and downloading GET /v1/videos/{id}/content) - the CLI holds no provider key, the gateway does, so renders show up in gateway logs, traces and pricing.

Disabled by default. While text_to_video.enabled is false, the TextToVideo tool definition is not sent to the LLM at all, so it costs zero prompt tokens. Avatar renders upload a picture of your face and a recording of your voice to a third-party provider - that is why this feature is opt-in rather than on by default.

Add a top-level text_to_video section to .infer/config.yaml (project) or ~/.infer/config.yaml (user):

text_to_video:
  enabled: true # feature flag (default: false) - tool absent from the LLM payload when false
  # model: elevenlabs/veo-3.1-fast-generate-001 # prompt renders
  # avatar_model: elevenlabs/creatify-aurora # lip-synced avatar renders
  # size: '' # "widthxheight", passed through to the provider; empty = provider default
  # output_dir: ~/.infer/projects/<project-slug>/tmp/media/video
  # timeout: 900 # seconds to wait for a render job
  # poll_interval: 5 # seconds between job status polls
  # create_avatar: false # also register the CreateAvatar agent tool
  require_approval: false # optional; unset = no approval, like the image tools

Gateway requirements: the Videos API must be enabled on the gateway (VIDEOS_ENABLED=true, gateway v0.54.0 or newer) and the gateway must hold credentials for the provider behind model / avatar_model (an ElevenLabs API key for the defaults). A gateway the CLI starts itself gets VIDEOS_ENABLED=true automatically while text_to_video.enabled is on, and an already-running instance without the Videos API is restarted. Point the CLI at an externally managed gateway and you set VIDEOS_ENABLED=true and the provider key there yourself.

All options live under text_to_video in .infer/config.yaml. Every key also has an INFER_TEXT_TO_VIDEO_-prefixed environment variable that takes precedence over the config file.

For example:

export INFER_TEXT_TO_VIDEO_ENABLED=true
export INFER_TEXT_TO_VIDEO_AVATAR_MODEL=elevenlabs/creatify-aurora

With text_to_video.enabled set, the agent gains a TextToVideo tool - see TextToVideo in the tools reference for the full parameter list. In chat, just ask:

• "render a 5 second clip of rain on a window" - prompt render with model.
• "make my alice avatar say this" (with a generated or recorded audio clip) - lip-synced render with avatar_model.

Limits: the portrait plus the audio clip travel in one request, so together they must stay under the gateway's 10 MiB request body limit - trim the clip or downscale the portrait if a render is rejected. creatify-aurora renders at 480p or 720p and keeps the portrait's aspect ratio, so a vertical portrait yields a vertical video.

Avatars live under ~/.infer/avatars/, one folder per avatar holding one or more portrait images (.png, .jpg, .jpeg, .webp) - for example a few angles of the same face:

~/.infer/avatars/
  alice/
    alice.png          # primary image, first in sort order
    alice-left.png
    alice-right.png
  bob/
    bob.jpg

Lip-sync models take a single image, so the CLI uses the first image in sort order within the folder. The library is userspace-wide (shared by every project) and is preserved by /reset along with the rest of your configuration.

How an avatar is sent depends on whether there is audio:

So a prompt render with a library avatar describes what the person looks like from several angles rather than pinning the opening shot, which keeps the face consistent across the clip. The gateway never combines the two: a render sends either a first frame or reference images, not both.

Requires gateway v0.55.0 or newer - the reference_images field does not exist on earlier gateways.

Per-model image limits are enforced by the gateway, which rejects an over-limit render with a 400 naming the limit:

Avatar (lip-sync) models such as creatify-aurora ignore reference images entirely. Keep folders small for Veo: the limit counts every image in the avatar folder.

# List avatars and their images
infer avatars list
infer avatars list --format json

# Create an avatar from a photo, generating extra views
infer avatars create alice --from ~/Pictures/alice.png

# Delete an avatar folder and its images
infer avatars delete alice

infer avatars create <name> --from <photo> copies the photo in as the primary image and then generates additional views of the same face - by default both three-quarter angles - through the gateway's image edit API using the tools.image_edit.model model.

Creating an avatar normalises the photo first: a JPEG is turned upright according to its EXIF orientation and re-encoded without its metadata, so a phone photo stored sideways arrives the right way up and camera and GPS tags never reach the library or a provider. The generated views come from that stored image, not the original file. PNG and WebP pass through unchanged.

create never overwrites an existing avatar - delete it first if you want to rebuild it.

Privacy: generating views uploads the photo to the image-edit provider (OpenAI by default), and an avatar render uploads the portrait and the audio clip to the video provider. Drop images into ~/.infer/avatars/<name>/ by hand to build a library without the image-edit round trip.

CreateAvatar is the agent-side twin of infer avatars create - it lets the model build an avatar from a portrait it just generated, a selfie sent over a channel, or a photo in the project, without you dropping to a shell.

It is opt-in on top of the feature flag: registered only when both text_to_video.enabled and text_to_video.create_avatar are true.

text_to_video:
  enabled: true
  create_avatar: true

export INFER_TEXT_TO_VIDEO_CREATE_AVATAR=true

It requires approval by default, unlike the other media tools, because it writes to your userspace avatar library and uploads a photo of a face. An explicit text_to_video.require_approval overrides that either way.

The tool never overwrites an avatar and has no delete counterpart - the agent can add to your library but not replace or remove anything in it. Photo normalisation (EXIF orientation, metadata stripped) applies here too.

• CLI - the tool reference entry, parameters, and the rest of the infer command-line tool
• CreateAvatar - the tool reference entry for building avatars from chat
• Videos API - how reference_images and input_reference differ on the wire
• Desktop App - manage the same avatar library from Settings -> Avatars, and place renders on a Content project timeline
• Text-to-Speech - generate the voice track an avatar render lip-syncs to
• TextToMusic and TextToSFX - the audio siblings, with the same output_path rules
• Configuration - full configuration system across the gateway and CLI