3D previz for AI video: block your scene before you generate, the complete method

The problem everyone knows: the generator cannot see your shot

A video generator understands a camera move when you show it, not when you write it. 3D previz makes what you need to show.

You have the shot in your head. A low tracking shot following a character down a corridor, the camera rising as he pushes the door, light pouring in. You write all of it in the prompt, you run it, and the model gives you a camera planted in the middle, a character walking toward the lens and a door that no longer exists. Second attempt, third, tenth. Each one costs between 20 and 60 credits, and you still have no guarantee the eleventh will be right.

It is not a vocabulary problem. Video models like Seedance 2.5, Kling V3 or Wan 3.0 read "tracking shot", "low angle" or "35 mm" perfectly well. The problem is that those words describe an infinity of possible shots, and the model picks one at random. What removes the ambiguity is a start frame, an end frame and, when the model accepts one, a reference video of the move.

That is exactly what a previz produces: the grey, simplified version of your shot, with real volumes, real mannequin characters and a real camera, that you set once and that serves as the reference for every generation after it. Cinema has done this for thirty years before shooting. AI video is just discovering it needs it even more.

What a previz is, and what it is not

A previz is not a finished 3D render: it is a blocking, simple volumes in the right place, filmed by a camera with the right values.

In traditional production, previsualization (previz to everyone) is the step where the shot is built in 3D before it is filmed. No textures, no actors, no detailed sets: boxes for walls, cylinders for columns, mannequins for the cast. What matters is elsewhere: where the camera is, which lens, at what height, which move, at what speed, and who stands where in the frame at every second.

Everything grey in a previz is grey on purpose. The day you add textures, you start discussing the texture instead of the framing. Previz exists to settle staging questions, and it settles them better when it does not look like the film yet.

For AI video, previz has one more job: it outputs files the generator can read. A clean image of the first frame, a clean image of the last, a few seconds of video where the move is exactly visible, and a description of the shot written in the terms the models understand. It is no longer a discussion tool, it is half of the prompt.

The four things a previz must give you

Start frame, end frame, reference video, structured prompt: with these four files, generation becomes execution, not a lottery.

The start frame fixes the framing of the first second. Every image-to-video model takes one: it is the guarantee that the shot begins exactly where you decided, with the right character in the right place. The end frame, accepted by Seedance, Kling and Wan, fixes the last second. In between, the model only has to interpolate a coherent move, which it does well.

The reference video is the grey render of the previz, as it is. Some models take it as a motion reference: Kling 3.0 Motion Control replays the gesture and trajectory of a video onto the character of an image. There, the camera you drew literally becomes the camera of the render. For the other models it serves as a visual reference and a check: you compare the render to the previz, shot by shot.

The structured prompt, finally, describes the shot in cinema terms: duration, ratio, starting position of the camera relative to the subject, height, lens and its feel (wide, portrait), move (a tracking shot of so many meters at such speed, a rise of so much, a pan of so many degrees), what the characters do, the mood and the lights. Written by hand, that paragraph takes twenty minutes and contains mistakes. Generated from the scene, it is exact, because it is computed.

All of it fits in a zip folder that Imaginode's 3D Previz studio exports in one click, with the camera metadata as JSON for those who want to go further. The studio guide details every file.

Doing your previz in the browser, without 3D software

One sentence is enough to get a scene, the mouse does the rest, and rendering is free: previz stops being reserved for people who can animate in Blender.

Until now, making a previz meant opening Blender, Unreal or Cinema 4D, modeling volumes, rigging a camera and learning to animate keys. Weeks of learning curve for someone who only wants to decide a framing. Imaginode's 3D Previz turns the problem around: you describe the scene in one sentence, with reference images if you have them, and an AI director builds it in 3D in your browser.

Concretely, "a hospital corridor at night, a nurse walking toward a door at the far end, low camera behind him" gives a corridor, doors, a named mannequin, cold light, a 35 mm camera placed 80 cm off the floor behind the character. It is not a generated image: it is an editable scene where every element can be selected, moved and adjusted. Building it costs from 5 credits, 12 on average (one credit is one cent), and the price is shown before you click, as everywhere on Imaginode.

Then you adjust. You select the camera, pick a lens from 14 to 135 mm, apply one of 25 moves (push in, crane, orbit, reveal, vertigo, handheld) or draw the path click by click on the floor. You record a virtual camera by hand in fly mode, then smooth it. You set the start frame and the end frame. You ask the AI director to "move the camera back two meters and switch to 50 mm" when you would rather talk than click.

And you export: image, mp4 video encoded right in the browser, contact sheet, storyboard of every shot, AI package. No credits spent on any of it. The bird and balloon previz that illustrates the studio page was rendered this way: 27 seconds of chase in one take, exported in 1080p in a few seconds.

From blocking to render: the path on the canvas

The previz lives in the same project as the generation: the exported frames become the inputs of a Video node, without going through your disk.

The full path is four steps. One, describe the scene and let the AI director block it. Two, set up the shot or shots: camera, lens, move, character positions on the timeline. Three, export the shot's AI package. Four, on the Imaginode canvas, plug the start frame and end frame into a Video node, paste the structured prompt, pick the model and generate.

For the model choice, the rule is simple. A shot where the camera moves a lot and the character has a precise gesture goes to Kling 3.0 Motion Control with the reference video. A shot where framing matters more than gesture goes to Seedance 2.5 with start and end frames, or to Wan 3.0 for long shots up to 30 seconds. Our guide on directing the camera of an AI video generator compares behaviors model by model.

One detail that changes everything: the previz is never redone. If the first render is not right, you switch models or adjust the prompt, but the scene, the camera and the frames stay. You no longer start from zero at every attempt, and that is where the credits are saved.

What it costs, compared with the "ten prompts" method

A previz costs a dozen credits once; every avoided video generation saves twenty to sixty.

Take a 10-second shot in 1080p. On the canvas, a Seedance 2.5 generation sits around 60 credits, Kling V3 a little below, Wan 3.0 above depending on resolution; the calculator gives today's exact values. Ten attempts to reach the intended framing is 400 to 600 credits, that is 4 to 6 euros, for a single shot.

With a previz, the scene costs a dozen credits, adjustments two or three per message, rendering and exports nothing. You then generate once or twice instead of ten times. On a twenty-shot film the gap is counted in tens of euros and in hours of waiting avoided. And the argument that matters most is not the price: it is that the final framing is the one you decided.

Higgsfield launched a prompt-driven 3D scene tool, 3D Jutsu, on September 4, 2026, on subscription. We looked at it closely in this article: the two approaches are alike in principle, they differ on price, integration with the generator and camera control.

Five previz mistakes that cost you later

A useful previz is plain, to scale, and designed for the model that will read it.

First mistake: wanting the previz to be beautiful. It has to be readable. Grey volumes and a mannequin are enough; a scene loaded with detail makes the camera move less visible on the reference video, and the model latches onto details that mean nothing.

Second: forgetting scale. A mannequin is 1.75 m, a door 2.10 m, a car 4.5 m. If proportions are wrong, the lens lies, and the render will look like a miniature. Imaginode's blocking kit places every object at its real size, and smart placement puts it on the floor.

Third: a camera that does everything at once. A tracking shot, a rise, a pan and a zoom in the same 8-second shot, no model follows. One main move per shot, two at most. The studio's 25 presets exist for that.

Fourth: neglecting the end frame. It is what keeps the model from drifting in the second half of the shot, and it is the first thing people forget to export. Fifth: not naming the characters. A mannequin named "Lea" in the scene becomes "Lea" in the structured prompt, and you can attach your reference character to it on the canvas.

Where to start

One shot, one sentence, one export: in ten minutes you will know whether the method suits you.

Pick a shot you have never managed to get through prompting. Open 3D Previz, describe it in one sentence, let the scene build. Adjust the camera with the mouse until the picture-in-picture shows the frame you had in mind, set the start frame and the end frame, export the package.

Then generate once on the canvas, with the right model. Compare with the grey render. If the shot is there, you just saved nine generations. If not, you know exactly what drifted, because you have a reference next to it. That is previz: a reference to judge against, instead of a memory of what you wanted.