How to turn a photo into a video

Every image-to-video model here takes the same two things — a still and a sentence about what should move — and they differ almost entirely in how long a clip they will give you, at what size, and out of which wallet. This page is the comparison, with the real envelopes rather than marketing ranges, so you can pick the lane before you spend on it.

Pick the lane before you pick the prompt

The single decision that matters is which wallet you are spending. Green credits (⚡) refill daily and buy only the models on our own GPUs. Gold credits (✦) are bought or granted by a plan and buy the hosted models. A prompt that works on one lane usually works on the other; a budget does not move between them.

The image-to-video lanes, with their real envelopes

Model	Runs on	Wallet	Clip length	Resolution	Plan
LTX 2.3 Image + Text → Video	Our GPUs	⚡ green	4, 8 or 16 s	up to 1080p	Every plan
MiniMax H3 Studio	Our GPUs	⚡ green	1-15 s	720p or 1080p	Every plan
Kling 3.0 Pro	Hosted	✦ gold	3-15 s	1080p	Starter and up
Google Veo 3.1	Hosted	✦ gold	4, 6 or 8 s	720p or 1080p	Starter and up
Seedance 2.0 Mini	Hosted	✦ gold	4-15 s	480p to 4K	Starter and up
Grok Imagine 1.5	Hosted	✦ gold	1-15 s	480p to 1080p	Starter and up

Which plan: A free account can run the two own-GPU lanes, with every render clamped to a 480 px short side. The four hosted models open at Starter and are billed in gold credits, which a free account has none of — the plan grants them, or you buy them.

Start frames and references are not the same input

Two different things get called image-to-video. A start frame is the literal first frame of the output: the model animates that exact picture. A reference is guidance — keep this face, this product, this style — and the model composes a new shot around it. Mixing them up is the most common reason a result looks nothing like the input.

• LTX 2.3 Image + Text → Video takes exactly one start frame. What you attach is frame one.
• Kling 3.0 Pro takes an optional start frame and an optional end frame, so it also covers guided transitions.
• Seedance 2.0 Mini takes an optional start and end frame, plus up to nine references (images, clips or audio) on the same endpoint.
• Google Veo 3.1 has three mutually exclusive lanes — plain text, a first frame with an optional last frame, or up to three Ingredients used as references.
• MiniMax H3 Studio takes no start frame at all. It takes up to nine reference items — nine images, three clips of 2-15 s, three audio tracks — and builds from them.
• Grok Imagine 1.5 takes an optional start frame and up to seven reference images.

Turning one photo into a clip

1. Open Generate on the video page. Go to /video. Generate is the default section.
2. Attach the still. Drop it on the composer's media row. The composer reads the image's aspect ratio and offers the canvases the chosen model can actually render at.
3. Pick a model for the length you need. This is where the envelopes above matter. If the shot needs sixteen seconds, LTX and its green wallet is the only lane that reaches it in one render. If it needs eight seconds of 1080p with sound, Kling 3.0 Pro or Veo does it in gold.
4. Write the motion. Name the camera move and the subject's action, in that order, and keep it to one shot. "Slow push in, she lifts the cup and looks off-frame left" is a brief. "Cinematic masterpiece, 8k" is not.
5. Read the quote, then submit. The Generate button carries the cost and the wallet it will come out of before you press it. Hosted lanes are priced per second, so length is the price.
6. Work on from the card. Every finished clip carries its own actions: regenerate with the full original inputs, reuse the inputs on another model, or hand it on to Motion or AI edit.

Open the composer — Opens the video composer with nothing selected. Attach your still and pick a lane.

What actually changes the result

The still fixes the subject, the framing and the light, which means three quarters of what people usually write into a prompt has already happened. The prompt's whole job is the fourth quarter: what moves, how fast, and where the camera goes.

• One shot per render. These models do not cut. Asking for "then it cuts to" produces a smear, not an edit. Generate two clips and join them.
• Name the camera separately from the subject. "Static camera, the flag moves" and "the camera orbits, the subject is still" are different renders and the model will not guess which you meant.
• Small motion reads better at short lengths. A four-second clip has time for one gesture. Sixteen seconds of LTX has time for a move plus a beat.
• Sound is a separate decision. Kling 3.0 Pro and Veo can generate their own audio, and it costs more when they do — Kling's sound option multiplies its per-second price.

Watch out: Wan 2.2, Wan InfiniteTalk and Wan Animate also run on our own GPUs, but they are payable in gold only unless the account is Ultra, and InfiniteTalk and Animate need Creator. They are not a cheaper route to green-paid video.

Aspect ratio, and why it is decided by the still

The composer reads the aspect ratio off the image you attach and offers the canvases the chosen model can actually render at. That is worth understanding rather than fighting: a 16:9 still forced into a 9:16 render has to be cropped or padded, and the model will make something up in the space it was given.

So generate the still at the ratio the finished video needs. If the destination is a vertical feed, make the picture 9:16 first. LTX offers 21:9, 16:9, 4:3, 1:1, 3:4 and 9:16; Veo publishes only 16:9 and 9:16; Kling 3.0 Pro renders its own native size whatever the picker shows. Fixing the ratio at the image step costs one cheap render and saves a re-shoot.

What goes wrong, and what to do about it

• Faces drift over the clip. Longer renders drift more. Cut the length, or fix both ends with a first-and-last-frame render instead of an open-ended one.
• Hands and text dissolve. Both are still the weak spot of every model here. Frame them out, or keep the motion small enough that they are not the focus.
• Nothing moves. The prompt described the picture instead of the motion. The still already supplied the picture; say what changes.
• Everything moves. Usually a prompt that named a camera move and a subject action and a background event. Ask for one.
• The clip is shorter than you asked for. Some lanes have a duration ceiling that depends on the canvas: a lane that renders sixteen seconds at 480p may cap lower at 1080p, and the composer refuses the combination rather than quietly trimming it.

One practical habit beats all of the above: test at four or five seconds on the cheapest lane that can plausibly do the shot, and only pay for length and resolution once the motion is right. A rejected 16-second 1080p render on a hosted model costs real money; the same idea proved out at 480p on the own-GPU lane costs a few green credits.

What it costs: Hosted lanes bill per second of output, so length is the whole price. Own-GPU LTX bills per started five-second block instead: its 4-second option is one block, 8 seconds is two, and 16 seconds is four. The composer quotes the real number on the button either way.

Which image-to-video model is free?

LTX 2.3 Image + Text → Video, and MiniMax H3 Studio for reference-driven shots. Both run on our own GPUs and are payable with the green wallet a free account refills daily, clamped to a 480 px short side on that plan.

What is the longest clip I can get from one photo?

Sixteen seconds on LTX 2.3, in one render, on the green wallet. Kling 3.0 Pro, Seedance 2.0 Mini and Grok Imagine 1.5 reach fifteen seconds in gold; Veo 3.1 stops at eight.

Can I control the last frame as well as the first?

Yes, on three lanes. LTX 2.3 First / Last Frame takes both images and is on the green wallet; Kling 3.0 Pro and Seedance 2.0 Mini take an optional end frame in gold.

Why does my clip drift away from the photo?

Usually because the lane you picked treats the image as a reference rather than a start frame. MiniMax H3 Studio and Veo's Ingredients lane are reference lanes by design. If you need the photo to be frame one, use LTX's Image + Text lane or attach the image as Kling's start frame.

Does a longer clip cost more?

Always. Hosted lanes bill per second of output, so the length is the invoice. LTX bills per started five-second block: 4 seconds is one block, 8 seconds is two, 16 seconds is four.