Niches · Video generation

Animate a still or keep a character across clips from reference images

Four or more companies' docs show a full fit.

4companies, full fit
12companies, fit with a limit
34companies carry the tag
21share of the category's work
11first-hand accounts

What counts as done

An MP4 of the requested length from the input image as first frame (or keyframes) and the subject's reference, from one create call and one poll loop, with the input accepted in the form the agent holds it (URL, base64 or upload).

Measured, for our own tests: the first frame (or keyframes) matches the input image and the subject keeps the reference identity. No docs we read state these figures, so they don't count against a listing here.

How agents ask for it: animate this image, turn this product shot into a five second clip, keep the same character in every scene, use this first and last frame

Does the job 4

The docs show it meets the test above, and nothing we read says otherwise.

ListingWhy
Pika API E
Mellis, Inc.
Image-to-video operations (Seedance 2.5, Wan 3.0, Pika 2.5 Keyframe) take image_url and end_image_url, with reference-to-video operations; a local file goes through POST /v1/media/uploads (20 MB image cap) or a public URL is passed straight in. source
Replicate video models B
Replicate, LLC
Kling 3.0 takes start_image and end_image (jpg or png, 10 MB, 300 px minimum) and Replicate accepts files as HTTP URLs or data URLs on every model; one create and one poll or webhook. source
Runway API C
Runway AI, Inc.
POST /v1/image_to_video takes promptImage with position first (and last on Veo 3.1 and Kling), as HTTPS URL, data URI or ephemeral upload; Kling O3 and Gemini Omni take up to 5 to 7 references. Inputs are auto-cropped to the ratio. source
Vidu API E
ShengShu AI HK Limited
Q3 image-to-video takes the start frame in images as URL or base64 data URI (png, jpeg, webp, 10 MB decoded), with separate start-end and reference-to-video endpoints, then one poll or callback_url. source

Does the job, with a limit 12

It does the job, but something in the docs or a user's account will get in an agent's way. The limit is named in each row.

ListingWhy
Alibaba Wan (Model Studio) B
Alibaba Cloud (Singapore) Private Limited
Model Studio docs couldn't be reached; the open Wan2.2 I2V-A14B and TI2V-5B weights animate a still at 720p, but A14B needs 80 GB VRAM on one GPU and a first-hand report saw two A100s run out of memory. source
fal video models BB
Features & Labels, Inc.
Kling v3 Pro takes start_image_url, end_image_url and element reference sets, but inputs are URLs per model schema, and a first-hand report shows Wan 2.2's end_image_url ignored except with the default prompt. source
framefields C
Gatewai
framefields animates stills as code layers, and the optional paid AI gateway can generate video from code, but the gateway's image-to-video models and inputs weren't documented in what was read. source
Google Veo BB
Google LLC
Veo 3.1 takes first and last frames and up to three subject images as base64 or Cloud Storage URIs, but first-hand reports show people refused by region or Child and Celebrity codes, base64 inputs failing, and the documented reference example returning 'use case not supported'. source
Higgsfield API B
Higgsfield Inc.
Kling 3.0 Pro image-to-video takes image_url as first frame and last_image_url as last, with elements; inputs must be public URLs or presigned uploads, and a first-hand report shows uploads failing every time with S3 SignatureDoesNotMatch. source
Luma AI API D
Luma AI, Inc.
Ray 3.2 takes start_frame, end_frame or up to 64 keyframes as URL, base64 data, file_id or a prior generation_id, but start and end frames only work at 5 seconds, and no face policy was found. source
Magic Hour MCP D
Magic Hour AI, Inc.
create_video_from_image takes image_file_path and end_image_file_path as a direct URL or an upload-URL file_path across Kling, Seedance, Veo, Wan and MiniMax, then wait_for_video_project, but base64 isn't accepted and support per model varies inside one schema. source
MiniMax Video API B
Nanonoble Pte. Ltd.
POST /v2/video_generation takes first_frame and last_frame images or up to 9 reference images as public URL or base64 data URI (30 MB, 256 to 5760 px), but a first-hand report shows reference-to-video adding wedding rings to every generation, via API and open weights. source
PixVerse API E
MOTIVAI PRIVATE LIMITED
Image-to-video, first and last frame, multi-transition and reference-to-video (Fusion) endpoints exist, but every image must first go through Upload Image (file or public URL, under 20 MB) to get an img_id; base64 isn't accepted inline. source
Popcraft MCP F
Popcraft
popcraft_generate_video covers image-to-video, first and last frame and reference-driven video with public-URL import or signed upload, but Seedance references of recognisable people are rejected, and the hosted tool schemas couldn't be read (the API path is disallowed by robots.txt). source
Seedance on BytePlus ModelArk C
BytePlus Pte. Ltd.
Takes first_frame, last_frame or up to 30 reference images as URL, base64 data URI or asset ID, but Seedance 2.5 and 2.0 reject reference images or video containing real human faces unless they are authorised assets or the vendor's own outputs. source
sync. labs MCP server D
Synchronicity Labs
Talking-photo workflows animate a portrait to speech, but that is lip sync rather than general image-to-video with keyframes; reference identity and input forms for it weren't read. source

Couldn't check 1

The docs were closed to us, or silent. These carry no mark either way.

ListingWhy
Kling AI API F
Kling AI Pte. Ltd.
The API reference renders only with JavaScript and llms.txt lists no request fields, so image input forms, end frames and reference counts couldn't be read.

Doesn't do this job 2

Listings in the category that do something else, with the reason.

ListingWhy
Clueso F
Desklamp, Inc.
Clueso places images and keyframed elements on clips; it has no model that animates a still into motion while keeping identity. source
HyperFrames E
HeyGen
HyperFrames renders HTML compositions; it doesn't generate motion from a still with a model (generated video comes through a separate HeyGen media skill). source

1 listings in the category were added after this reading and aren't marked yet.

What stops agents finishing

  • reference image of a realistic or AI-generated person refused by region or face policy with no appeal route
  • documented reference-image mode returns 'use case not supported' for the key
  • input upload fails or needs a hosted URI the agent does not have
  • open-weight image-to-video needs more GPU memory than the host has

Around it

Before: image-generation.text-to-image, image-generation.character-consistency. After: video-generation.extend-to-length, storage.upload.

Where this comes from

Weight 21 of 100 in Video generation: 11 of 37 accounts; the job with the most failure reports (Veo reference images, face policy blocks, Higgsfield uploads, Wan end frames, MiniMax reference drift)

The same as data: /api/v1/use-cases/video-generation.image-to-video.json, or the MCP tool get_use_case with id video-generation.image-to-video. Other jobs in this category: /api/v1/niches?category=video-generation.

Draft, read 2026-11-15 from public docs and first-hand accounts, not yet reviewed by a person. A mark says what the docs show, not what a run measured.

For companies

Do agents find, use and choose your tools?

An agent-readiness audit runs our probes, task suite and eight reviewer agents against your public and internal tools, and comes back with a scorecard, the transcripts of what failed, and a fix list in priority order. From $2,500, re-run included. We never take payment to move a rank. We do help companies earn one.