vidmoat-image-live-1

Model card for vidmoat-image-live-1: bring a still to life without adding anything to it. Tiers, inputs, output, prices, speed, safety checks and known limits.

vidmoat-image-live-1 moves what a picture already shows. Give it a still and a sentence ("slow push-in", "the waterfall flows", "she breathes and blinks") and it returns a short clip of that picture in motion. It is built for product shots, thumbnails, slides, memories and b-roll: the cases where a full text-to-video model would reinvent the scene.

What it does#

A router reads the request and picks the cheapest tier that can do it. A quote always tells you which one you will get.

Tier	Motion	Engine	Price	Length
Camera	Push in, pull out, drift, orbit, pan, with real depth parallax	Depth estimation and parallax rendering	8 credits a second	2 to 10 s
Cinemagraph	One region moves, everything else stays still: water, waves, clouds, smoke, steam, fire, leaves, grass, hair, flags	Masked flow on the region	8 credits a second	2 to 10 s, can loop
Object	A cut-out subject floats, bobs, sways or tilts	Subject cut-out and motion	8 credits a second	2 to 10 s
Generative	Faces breathing or blinking, people moving, vehicles driving, turning objects, any motion described in words	Our own deployment of Wan 2.2 TI2V-5B (Apache 2.0) on our GPU	24 credits a second	4 or 5 s, no loop

Setting move, region, motion, box, strength or direction keeps a request on the first three tiers, which implement those controls exactly. Leave them out and describe the motion in words to reach the generative tier.

Output#

MP4 (H.264), 24 frames a second, 720p on the short side, no audio track.
Aspect: the source's, or 16:9, 9:16, 1:1 (and 4:5, 5:4, 4:3, 3:4 on the first three tiers).
Delivered at a URL in your account, ready to use as an addClip source in a project.

Price#

You pay for exactly the seconds you ask for, at the tier's rate: a 5-second camera move is 40 credits, a 5-second generative clip is 120. quoteOnly: true returns the price without charging, and maxCredits refuses anything above your ceiling before it starts. A clip that fails its checks is refunded in full, automatically.

Speed#

Tier	Typical time to a finished clip
Camera, cinemagraph, object	Under a minute
Generative	About 2 to 5 minutes. The GPU starts for each burst of work, so the first clip after a quiet spell takes longest.

One living image runs at a time per account; a second request while one is running gets 402 quota_exceeded.

What it will not do#

Add things. Anything appearing, pouring, breaking or transforming is refused before charging with a 400 that points to Generate a video, which is built for that.
Sound. Clips are silent; add music or speech in the project.
Long clips. 10 seconds on the first three tiers, 5 on the generative one.
Two kinds of motion at once. A camera move and ambient motion together are refused today rather than half done.

How each clip is checked#

Every clip is measured before you get it, and one that fails is not delivered and is refunded:

Nothing new appears. On the generative tier a detector counts the people, animals and vehicles in your still and in the clip. If one appears that the still does not have (the base model invents a subject about 1 time in 7), the clip is rejected, retried once with a new seed, then failed and refunded.
It actually moves, and on the cinemagraph tier only inside its region.
No cuts and no drift away from your picture.
Prompts pass the same safety screen as every Vidmoat generation; a refused prompt is a 422 prompt_rejected and costs nothing.

Known limits#

Faces are animated gently; expressions and speech are out of scope.
The generative tier has a monthly GPU allowance. When it is used up, generative requests get 503 unavailable until the 1st and are not charged, while camera moves and cinemagraphs keep working.

Calling it#

curl

curl -X POST https://api.vidmoat.com/v1/ai/live-images \
  -H "Authorization: Bearer $VIDMOAT_KEY" \
  -H "Content-Type: application/json" \
  -d '{"image":{"url":"https://example.com/portrait.jpg"},"prompt":"she breathes and blinks, hair moves in the breeze","durationSec":5,"aspectRatio":"9:16"}'

# then, every five seconds:
curl https://api.vidmoat.com/v1/ai/jobs/$JOB_ID -H "Authorization: Bearer $VIDMOAT_KEY"

Every parameter and error is in the API reference. Over MCP the same model is the generate_live_image tool, polled with get_generation_status.

Model details#

Name	`vidmoat-image-live-1`
Released	October 2026
Generative base	Wan 2.2 TI2V-5B (Apache 2.0), run by Vidmoat at 12 steps and upscaled to 720p
Other tiers	Vidmoat's depth parallax, masked cinemagraph and object motion pipelines
Weights	Not published. Available through the API, MCP and the Vidmoat apps.