Image to video engines compared — which one animates your still

Handing an engine a still is the most controllable way to make a video. The subject, the framing and the light are already decided — they are in the picture — and the engine has one job left, which is to move it. Every video engine on eroq accepts a start frame, so the question is never whether an engine can do image-to-video. It is what each one does after the first frame.

What a start frame actually buys you

In text-to-video the engine invents the subject, the composition, the light and the motion all at once, and a flaw anywhere ruins the take. In image-to-video three of those four are already locked — a face you approved, a product photographed the way the client signed off, a painting whose palette took four renders to get right.

The workflow follows from that. Draft the frame in the Image studio until it is right, then animate it. Every render is saved in your Library with its recipe, and Animate on an image card sends it into the video tool as a source. A Character works differently: @mention it and engines that read reference pictures (Seedance, Kling, Veo 3.1, Hailuo 3, Wan) carry its face into the clip.

The roster, judged on what happens after frame one

Motion One is eroq's own uncensored engine, and that settles a whole category of jobs by itself. It and its sibling Motion Pro are the only video engines not moderated upstream, so when a provider refuses a still — a dark, violent or intense scene — they are the alternative. Motion One is on every account, free included, at 29 credits / second (145 credits for 5 seconds), 5 to 15 seconds at 480p or 768p (1080p and 2K upscaled), with sound baked in, and it takes a first frame, a last frame or both. Motion Pro renders 720p or 1080p from a start frame. The rest of this article covers the third-party engines, which are moderated upstream by their providers.

Kling 3.0 is the one to hand a face. Its reputation is fluid motion and faces, which in practice means a portrait animated on Kling tends to stay the same person from the first frame to the last rather than sliding into a cousin halfway through. 720p, 3 to 15 seconds, sound, 270 credits for 5 seconds, Hobby plan and up (Kling 3.0 Pro on Creator). If your still is a person and the clip is about the person, start here. There is a side-by-side with Seedance on /compare/seedance-vs-kling.

Hailuo 3 Max is the one to hand a scene with weather in it. Its reputation is cinematic physics and dramatic light — smoke, water, falling debris, hard shafts of light doing something over time. If what should move in your still is not a person but the air around them, Hailuo is the pick. 480p or 768p, 5 to 15 seconds, silent, 170 credits for 5 seconds, Hobby and up; Hailuo 3 on Creator does the same at 2K with sound.

Seedance 2.5 is the one with the controls and the length. It honors a reproducibility seed, so once a take works you can re-render variations from the same starting point instead of re-rolling the dice, and it supports first and last frame interpolation — give it the still you start on and the still you want to land on, and it fills the middle. 4 to 30 seconds, 480p or 720p (1080p and 2K upscaled), sound, Creator and up, 490 credits for 5 seconds. One limit: the Seedance engines turn down a start frame that looks like a photo of a real person; illustrations and 3D renders pass.

Seedance 2.0, 2.0 Fast and 2.0 Mini are the rest of the family, 4 to 15 seconds, all seeded, all with both frames. Mini, on every account, at 60 credits for 4 seconds, is the family's cheapest way to learn whether a still animates well before you spend more on it.

The Veo 3.1 family animates a still and renders sound with it — ambience, effects, dialogue — at 4, 6 or 8 seconds, up to 1080p on Lite and 4K on the other two (4K from Creator): Veo 3.1 Lite on every account, Veo 3.1 Fast from Hobby (344 credits for 8 seconds), Veo 3.1 on Creator. It takes an end frame and a seed too.

Write the prompt for motion only

This is the single biggest mistake in image-to-video, and it is easy to fix. When a still is already the subject, re-describing the subject does nothing useful and can actively fight the frame you gave it. The engine can see the red coat. Telling it about the red coat only gives it an opportunity to render a different one.

So strip the prompt down to the things that change over time: what moves, how the camera moves, what the light does.

Say your source still is a woman standing at a rain-streaked window at night, made in the Image studio and already exactly right. The prompt is not a description of her. It is this:

She breathes out slowly and the glass fogs at the edge of the frame, one hand lifting to wipe a short arc clear; rain keeps running down the outside of the pane and the neon behind it pulses across the water. Slow push-in, a few degrees of handheld drift, everything else holds still.

Every clause is a verb or a camera instruction. "Everything else holds still" is doing real work — it tells the engine that the rest of the frame is finished, which is the whole premise of animating a still.

Compare that to the usual first attempt — "A beautiful woman in a red coat at a rainy window at night, neon lights, cinematic, 35mm". That is a text-to-video prompt pasted into an image-to-video job. It describes what the picture already is, asks for no motion at all, and comes back looking like a photograph with a mild breeze in it.

A three-line checklist for a motion prompt:

1. One action. She wipes the glass. The car door opens. The smoke lifts.
2. One camera move. Push-in, orbit, slow pan, handheld drift. Named, not implied.
3. One line about what the light or the environment does — or an explicit note that it holds.

The end-frame trick

If you know where the shot has to land — a logo resolving, a face turning to camera, a door fully open — many engines take both ends: the Seedance, Veo, Kling and Hailuo families and Motion One. Render two stills in the Image studio, set the first as the start frame and the second as the end frame, and write a prompt that describes only the journey between them.

This is the closest thing on the roster to directing rather than prompting, and it is what makes a sequence cut together — shot two can start on the exact frame shot one ended on.

What image-to-video does not do

Worth saying plainly. There is no lip-sync on eroq — an animated still will not mouth a voice track you bring; engines with sound perform the lines written in the prompt instead. Upscaling is its own step: eroq Upscale (coming soon) takes a finished clip to 1080p or 2K, from Video › Edit › Upscale or the Upscale button on a take. Changing who is in a finished clip is a job for Replace, not image-to-video. A clip comes back at the resolution and length you asked for; the trims, crossfades, titles and voice track come afterwards, in the Editor or in Premiere or After Effects if that is where you already work.

The short version

• A provider refuses the still → Motion One or Motion Pro, eroq's uncensored engines.
• The still is a person and the clip is about the person → Kling 3.0.
• The still is a scene and the motion is physics, weather or light → Hailuo 3 Max (or Hailuo 3 at 2K).
• You know the last frame, or you need a seed → Seedance (Mini to test the pair, 2.0 or 2.5 to finish) or Veo 3.1; longer than 15 seconds → Seedance 2.5.
• It has to arrive with sound → most engines do; Veo 3.1 Fast is the usual pick.
• You are not sure the still animates at all → Seedance 2.0 Mini, 4 seconds, 60 credits, then commit.

FAQ

Can every eroq video engine animate an image?

Yes. Every engine accepts a start frame, whether it comes from an upload or a render in your Library. What differs is length, resolution, plan access and what each engine is good at once the frame starts moving. The Seedance engines refuse start frames that look like a photo of a real person.

Which engine keeps a face consistent in image-to-video?

Kling 3.0, which is known for fluid motion and faces. Using a Character rather than a loose photo helps further: on engines that read reference pictures, its pictures carry the same face across both images and clips.

Which engine supports a start and an end frame?

The Seedance, Veo 3.1, Kling 3.0 and Hailuo families and Motion One all take a first and a last frame, so you can give the engine the frame to open on and the frame to land on and let it fill the middle. Seedance 2.0 Mini is the cheap way to test the pair; Seedance 2.5 goes longest.

Draft the frame, then write three lines of motion — open the video studio.