@antseed/provider-venice-video

Relays native Venice video requests to the Venice API.

Configuration

antseed plugin add @antseed/provider-venice-video
export VENICE_VIDEO_API_KEY=<seller-api-key>

• VENICE_VIDEO_API_KEY: Venice API key, sent as a bearer token
• VENICE_VIDEO_BASE_URL: optional, defaults to https://api.venice.ai
• ANTSEED_ALLOWED_SERVICES: Venice video model names
• ANTSEED_SERVICE_UNIT_BILLING_MODELS_JSON: venice-video pricing using video_generations or video_seconds
• ANTSEED_SERVICE_CAPABILITIES_JSON: optional video options to advertise
• VENICE_VIDEO_MIN_BALANCE_USD: optional, default 0. Models stop being advertised while the key's Venice balance (USD + DIEM + bundled credits) is at or below it

Each seller health sweep runs two free checks. The account check reads GET /api/v1/api_keys/rate_limits (key, access, balance): a rejected key, blocked access, or a balance at the minimum stops every model being advertised. The model check reads Venice's public GET /api/v1/models?type=video: a model Venice no longer lists stops being advertised. The list doesn't count against the key's video request limit. A model is advertised again once both checks pass. Neither call generates media.

Venice's model list does not say which image-to-video models accept a last frame, so the seller asks Venice at startup. For each image-to-video model without configured video.inputs, it sends one create that Venice must reject (empty prompt, impossible duration). Venice lists every problem in its 400 answer and names end_image_url when the model can't take one, so the check starts no job and costs nothing. Models Venice doesn't object to announce end_image_url. The checks are paced to stay under Venice's 40 video requests a minute; an unclear answer announces the first frame only. To skip the check, or to require frames, announce them yourself:

{
  "seedance-2-0-image-to-video-basic": {
    "video": { "inputs": ["first_frame", "last_frame"], "requiredInputs": ["first_frame"] }
  },
  "pixverse-c1-transition": {
    "video": { "inputs": ["first_frame", "last_frame"], "requiredInputs": ["first_frame", "last_frame"] }
  }
}

With antseed seller start, set this per service as seller.providers.<name>.services.<id>.capabilities.

At startup the seller reads Venice's video model list (GET /api/v1/models?type=video) and announces the queue fields each model takes as supportedParameters: prompt, duration, resolution and aspect_ratio when the model lists values, audio when it is configurable, and the frame fields for the video.inputs you announce (image_url, end_image_url, reference_image_urls). When you announce none, image-to-video models get image_url, plus end_image_url when the startup check says the model takes a last frame. Buyers send only the fields listed. A service that sets supportedParameters itself keeps it. If the list cannot be read or does not include the model, the service is announced without it.

Aspect ratio and audio

Venice prices a clip by its aspect ratio and audio as well as its resolution and duration (for example Seedance 2.5 at 480p for 4 s: $0.51 at 16:9, $0.67 at 21:9; Veo 3.1 costs less without audio). Price them with components that match aspect_ratio and audio ("true" / "false") next to resolution (metadata v16, up to 64 components per model), for example:

{ "unit": "video_seconds", "priceUsd": 0.0701, "match": { "resolution": "480p", "aspect_ratio": "16:9", "audio": "true" } }

Buyers price a create that leaves out aspect_ratio at the first announced one, and one that leaves out audio at video.audio, and the plugin sends Venice the same:

• video.aspectRatios: a create with another aspect_ratio is answered 422 unsupported_video_option before payment and without calling Venice. A create without one is sent with the first listed.
• video.audio: false (a silent service) refuses audio: true and, for a model whose audio Venice lets you turn off (audio_configurable), sends audio: false when the create leaves it out (Venice adds audio by default). While Venice's model list hasn't been read (or lacks the model), a create without audio is refused instead, so it must say audio: false.

A service with video.audio: true accepts audio: false; give it its own price if Venice charges less. Options you don't announce are relayed as sent.

Video edits

Models Venice marks video_input (for example minimax-h3-max-recast, kling-v3-standard-motion-control, wan-2-7-video-to-video) take a source video as video_url. For each one, the seller announces video and reference_image in video.inputs, and video_url and image_url (the character for motion control; Venice accepts it on every edit model) in supportedParameters. To learn whether the model takes reference images, it sends one create Venice must reject (empty reference_image_urls, impossible duration). Venice answers that a reference is required, not supported, or says nothing, and nothing is queued. A required one is announced in video.requiredInputs, and an accepted one adds reference_image_urls. Announcing video.inputs yourself skips this check. The seller also announces each edit model's resolutions from Venice's model list, in Venice's case.

The buyer pays for duration. Venice bills an edit listing only Auto durations for the whole source, and an output shorter than duration is never charged, so for every edit model the seller measures the source and answers 422 invalid_video_duration, before payment and without calling Venice, unless duration is the source length in whole seconds, rounded down (for example "6s" for a 6.2 s source). The source must be inline (data:video/mp4;base64,...). Otherwise the seller does not check the source; Venice does. A job Venice rejects after it was queued (for example motion control or Recast on a video without a person, or a source that is too short) is answered on retrieve with a JSON 400/422. The plugin turns that into {"status":"FAILED","error":"<reason>"}, so the job's payment channel is closed at once instead of when it expires. Venice answers a body above 32.5 MB with an HTML 413; the plugin turns that into a JSON 413 request_too_large.

The plugin relays POST /api/v1/video/queue unchanged, apart from the announced options above, and handles POST /api/v1/video/retrieve by returning Venice JSON status or streaming the finished MP4 through the seller. For private models, Venice returns a short-lived download_url on queue; the plugin removes it from the buyer's response, keeps it for 24 hours, and streams the file from it once retrieve reports COMPLETED. Kept URLs are saved in the seller's data directory (venice-video-download-urls.json, set through ANTSEED_DATA_DIR, which antseed seller start passes automatically), so they survive a seller restart. Other Venice model metadata is not enforced. POST /api/v1/video/complete is not relayed.

Check that Venice's terms allow your offering. AntSeed sellers must add value rather than resell raw API access.