EditorNodesPricingBlog

AI VIDEO GENERATOR

Six ways to make a shot

Generate a clip from a written description, animate a still you already have, morph between two frames, sketch the motion you want, raise the resolution of footage you shot, or move the performance from one clip onto a different character. All of it runs in the browser.

Open the video workspace

Updated:

One workspace, one type dropdown

The video workspace is a single surface with a generation type dropdown at the top holding five entries. Image to Video, Text to Video, Start-End Frame Video, Draw to Video and Video Enhancer. You pick the type, the form below rebuilds itself with the controls that type needs, and you generate.

The form also changes with the model, and on video that matters more than it does on images. Duration ranges, available resolutions, audio generation and the negative prompt field are all model properties rather than workspace properties. The model picker prints the resolution range and the duration range next to each name, so you can read what a model will accept before you select it.

Results from every type land in the same place. The Artifact Gallery on the My Creations page holds them, and any completed task can be picked as the input to a later generation through the Previous Task option. That is what makes chaining these tools practical rather than theoretical.

The six ways in

Each has its own page covering the models it offers, the settings it exposes and the work it is genuinely good at.

Text to Video

Write a prompt and generate a clip with no input footage. Pick the model, duration, aspect ratio and resolution.

Image to Video

Animate a still you already have. Supply the first frame, optionally the last frame, and describe how the shot should move.

Start-End Frame

Give the model where the shot begins and where it ends, and it generates the transition between the two images.

Video Upscaler

Raise the resolution of footage you already have to 1080p, 2K or 4K without regenerating the content.

Motion Control

Take the motion from a reference clip and apply it to a different character, or drive a camera move from a still.

Draw to Video

Sketch arrows and notes over an image on the built in canvas, then let the model read the drawing as motion direction.

Choosing between generating and animating

The first real decision is whether the shot should start from a prompt or from a frame. Text to Video gives the model everything to invent, which is the right call when the shot does not exist yet and you are still exploring what it should look like. It is also the least controllable, because composition, wardrobe, lighting and framing are all being decided at once.

Image to Video removes most of that uncertainty. The first frame is already settled, so the prompt only has to describe movement. Generating a still first, in the image workspace, and animating it second gives you a cheap approval step in the middle. You can look at the frame and reject it before spending a video generation on it.

Start-End Frame Video goes one step further and fixes both ends of the shot. Where you know what the audience sees at the first frame and the last, and the only open question is the movement between them, giving the model both images removes the ambiguity entirely. Reveals, morphs and before and after comparisons are the obvious cases.

A working sequence

Most finished work passes through more than one of these. A common run starts in the image workspace to establish the frame, moves to Image to Video to put it in motion at a short duration and a low resolution, and repeats that pair until the movement reads correctly. Only then is it worth generating the same shot long and at full resolution.

Testing short is the single largest saving available here. Video models charge by the second, so a 5 second test of a prompt that turns out to be wrong costs a third of what the same mistake costs at 15 seconds. Duration is the last thing to raise, not the first.

Delivery is the final pass. The Video Upscaler raises resolution once the content is locked, which is the right order, because upscaling a clip you then decide to regenerate is wasted work. Motion Control sits outside this line entirely. It takes footage that already carries the performance you want and moves that performance onto a different character.

Reading the model picker

The model dropdown in the video workspace prints two figures beside every name, the resolution ceiling and the duration range. Those are the most useful numbers in the interface, because both constrain controls further down the form. Selecting a 720p model and then hunting for a 1080p option is wasted effort, since that option was removed the moment the model was chosen.

The ranges also differ in kind. Some models expose a continuous slider, so any length between the end points is available. Others accept only specific lengths, which is why certain Seedance variants read 5s, 10s and 15s rather than a span. A plan that assumes 7 seconds is available will fail on the second kind.

Limits are set per generation type rather than per model name, so the same model can behave differently in two forms. Seedance 2.0 Mini reaches 4K in image to video and stops at 1080p in text to video. Read the figures in the form you are actually using rather than remembering them from the last one.

The Video Enhancer holds two different jobs

Four of the five generation types do one thing each. Video Enhancer is the exception. It opens on a second dropdown offering Video Upscaling and Motion Control, and those two share a form while sharing almost nothing else.

Upscaling raises resolution with the ByteDance Video Upscaler and changes nothing about the content. Motion Control uses Kling 3.0 Pro to re-animate footage from a reference character image, which changes who is on screen entirely. Different models, different required inputs, different results.

The form opens on upscaling, so motion control has to be selected deliberately. That is the step people miss, and it is why the two have separate pages here rather than one.

Where the cost actually goes

Video billing works differently from image billing, and the difference changes how you should work. Most video models charge per second of output, so duration is a direct multiplier rather than a preference. Higher resolutions cost more on top of that, which means duration and resolution compound.

The practical consequence is that the order of operations matters more here than in the image workspace. Finding out that a prompt produces the wrong movement costs a third as much at 5 seconds as it does at 15, and less again at the bottom of a resolution range. Every question worth asking should be asked at the cheapest settings that can answer it.

The exact figure appears in the form before you submit, and it updates as you change the model, the duration and the resolution. Failed tasks are refunded automatically during credit reconciliation, so a generation that errors does not cost you anything.

Model tutorials on the blog

Longer coverage of individual models, with sample output and the cases where each one struggles.

Luma Ray 3.2

Motion quality and prompt handling across the Ray 3.2 duration options.

Gemini Omni Flash

The fast default in the video roster, and where its 720p ceiling starts to show.

Seedance 2.0

How the Seedance family behaves on camera movement and scene coherence.

Kling 3.0

Negative prompt handling and audio generation on the Kling video models.

Vidu Q3 Drama and Turbo

The longest single generation in the roster, and the fast variant beside it.

LTX 2.3

Where LTX fits against the rest of the roster on speed and motion fidelity.

Common questions

What is the difference between the video generation types?

Text to Video builds a clip from a written description with no input file. Image to Video animates a still you supply. Start-End Frame Video takes two images and generates the transition between them. Draw to Video reads arrows and notes you sketch over an image as motion instructions. Video Enhancer works on footage that already exists, either raising its resolution or re-animating it from a reference character image.

No. Every generation type runs in the browser. There is no model download, no local inference requirement and no dependency on the hardware in your machine. You sign in and generate from the video workspace.

Typically 30 seconds to 5 minutes. The time scales with the model, the duration you asked for and the output resolution, so a 15 second 1080p clip takes considerably longer than a 5 second 720p test.

The maximum comes from the model rather than the workspace, and the model picker shows the range next to each name before you select it. Most models sit between 3 and 15 seconds. Vidu Q3 Drama reaches 30 seconds, which is the longest single generation in the roster, and Vidu Q3 Pro accepts durations as short as 1 second.

Each generation uses credits. Most video models charge per second of output and higher resolutions cost more, so the figure changes with the model, the duration and the resolution you picked. The exact cost appears in the form before you submit. Failed tasks are refunded automatically during credit reconciliation.

Yes, and most finished work does. Any completed task can be selected as the input to a later generation through the Previous Task option, so generating a frame, animating it and then upscaling the result is a normal sequence. The Nodes Graph Editor does the same thing visually and runs the chain in one pass.

Results appear in the Artifact Gallery on the My Creations page, reachable from the Creations menu on the right side of the workspace. You can download each result from its Artifact card, and everything is also written to your cloud storage folder.

Commercial use depends on the licence of the model you generated with, and those differ across the roster. Check the terms and conditions and the coverage of the individual model on the blog before publishing work from any single one.

    © 2026 - AI FILMS LLC