Yes — a free plan lets you compose multi-scene motion graphics with voiceover and captions. Paid plans (Creator $29.90/mo, Pro $99.90/mo) add AI image, video, and sound generation and remove the watermark.

Can I edit the video after it's generated?

Yes. Onda produces an editable scene graph, so you can change any element — text, color, timing, layout — without re-generating the whole video.

Do I need video-editing or design skills?

No. You describe the video in plain language and an AI director composes it — there is no timeline to learn.

A finished MP4 in any aspect ratio — vertical, square, or widescreen — for social, ads, and product videos.

Is there a watermark?

Free exports include a small watermark; Creator and Pro plans remove it.

The regeneration spiral: why you can't edit AI video — and the structural fix

The regeneration spiral is the loop you fall into when you try to fix one detail in an AI-generated video and re-rolling the prompt quietly changes everything else. You ask for a different headline; the camera move shifts, the colors drift, the timing breaks. You are now further from the clip you liked than when you started — and you have spent more credits to get there.

This isn't a prompting skill issue. It's structural. Below is why one-shot AI video can't be edited, and the fix: treating a video as a deterministic scene graph instead of a bag of pixels.

Why "almost right" is the worst place to be

One-shot text-to-video generators are remarkable at the first 90%. You describe a scene, and seconds later you have something that looks close to what you imagined. The problem is the last 10% — the part that makes a clip actually usable.

A real deliverable has hard constraints. The product name has to be spelled correctly. The brand color has to be exact. The voiceover has to say the approved line, not a paraphrase. The logo has to be the current one. "Close" doesn't ship.

So you go back to fix the one wrong thing. And here the trouble starts.

Pixels have no parts

A one-shot generator doesn't produce a video made of editable objects. It produces frames — a field of pixels predicted all at once from your prompt and a random seed. There is no "headline layer," no "background object," no "the part where the logo sits." There is only the finished image, baked.

When you change the prompt to fix the headline, you don't reach into the video and edit the headline. You generate a new video from a slightly different prompt. Everything is recomputed. The model has no obligation to keep the 90% you liked, because it never knew which 90% you liked — to the model it was all one inseparable output.

That is the regeneration spiral:

You re-roll to fix one thing.
The rest drifts, because nothing pins it in place.
You re-roll again to recover what drifted.
Costs and time climb while quality wanders sideways.

The spiral isn't a bug in any one product. It's the inevitable consequence of representing a video as pixels with no internal structure.

The fix is structure, not a better model

A bigger or smarter generator doesn't escape the spiral — it just produces a more convincing 90% that still can't be surgically edited. The escape is a different representation of the video itself.

Instead of predicting pixels, you produce a scene graph: an explicit, named description of every element and how it moves. A headline is an object with text, a font, a position, and an animation. The background is its own object. The voiceover is a tracked line of copy. The logo is a referenced asset. The whole video is data describing a composition, not a finished image.

From that scene graph, a render engine deterministically produces the MP4. Same graph in, same frames out — every time.

Two properties fall out of this, and they are exactly what editing requires:

Addressability. Every element has a name, so you can point at one thing.
Determinism. Rendering the graph is reproducible, so touching one element doesn't disturb the others.

Together they give you the thing one-shot generation can't: change one word and only that changes.

What that feels like in practice

In Onda Studio, you describe a scene and an AI director composes it — motion graphics, AI imagery, AI video, AI voiceover, and sound — into a deterministic scene graph. A source-available, GPU-native engine renders it to a real, editable MP4.

When you want to fix the headline, you fix the headline. The camera move stays. The colors stay. The timing stays. Refinement is an edit to one object, not a fresh roll of the dice. You can do it by chatting with the agent or by manipulating the element directly — both reach the same underlying graph.

There is no spiral, because there is nothing to spiral. The 90% you liked isn't at risk every time you touch the 10% you didn't.

Where one-shot generation still wins

Structure is the right answer for motion graphics — explainers, launches, kinetic typography, product demos, anything that is text, brand, and designed motion. It is not the right tool for photoreal footage of a place that doesn't exist, a dreamlike scene, or live-action B-roll. For those, dedicated one-shot generators like Runway, Sora, Veo, Kling, and Luma lead on raw fidelity, and the regeneration spiral is a cost you accept for capability you can't get any other way. (Onda can generate video clips too, for when you want footage inside an otherwise editable scene — it just isn't trying to win the pure-photoreal contest.)

The mistake is reaching for a one-shot pixel generator when what you actually need is an editable, on-brand motion-graphics video. That's the job a scene graph is built for.

The takeaway

The regeneration spiral is not a prompting problem you can outrun with practice. It's what happens when a video has no editable parts. The structural fix is to represent the video as a deterministic, addressable scene graph so that editing one element leaves the rest untouched.

If you've been re-rolling prompts hoping the next one finally lands, that's the spiral talking. You can describe a scene and then actually edit it at studio.onda.video.