H3 Max 3D to Video: Your Blockout Is the Prompt

H3 Max 3D to Video turns a rendered 3D blockout into a finished video. The source clip guides the layout, camera, object motion, edit timing, and object count. A short Scene Intent explains ambiguous proxies. Optional reference images guide how the subjects, materials, and environment should look.

That division of labor matters. This is not a prompt that happens to include a Blender render. The render is the directing layer. If a camera move or cut is wrong in the blockout, a more elaborate prompt is unlikely to repair it.

Three inputs, three different jobs

The cleanest way to plan a 3D-to-video job is to decide which input owns each decision.

The official API schema accepts a required video_url, an optional prompt, optional reference_image_urls, a limit for automatically generated references, and an output resolution. The input video can be up to 15 seconds and contain up to 32 shots. The available output resolutions are 480P, 768P, and 1080P, and the output duration follows the source clip.

Think of it as a handoff between departments:

Blockout = where and when
Scene Intent = what the proxies mean
References = how the scene should look
Output = a finished video, not an editable 3D scene

Make the blockout readable before making it beautiful

Simple geometry is fine. Ambiguous geometry is not.

A gray cylinder can stand in for a person if it is clearly separated from the set and its movement is easy to follow. Five similar cylinders crossing each other, passing behind a wall, and changing direction during a whip pan give the model a much harder problem. Extra polygons do not solve that. Better staging does.

Before exporting the preview, check five things:

1. Silhouettes stay distinct. Important subjects should not merge into the background or each other.
2. The camera already tells the story. Set the move, lens feel, framing, and cuts in the 3D scene rather than describing a replacement move in prose.
3. Motion has a clear start and finish. Avoid a subject beginning or ending a critical action exactly on a cut.
4. Proxy roles remain consistent. A block that means “delivery van” in shot one should not become an unrelated prop in shot two.
5. The edit earns every shot. More cuts can increase both ambiguity and cost, even when the total timeline stays short.

A Workbench, EEVEE, viewport, or playblast-style render can carry this information. The endpoint receives a video, not a .blend, Maya, Cinema 4D, Unreal, or SketchUp project. Keep the editable scene file because the returned result is a finished video—not a mesh, rig, material graph, texture set, or camera file.

Scene Intent should label the scene, not rewrite it

The API describes the optional prompt as a clarification of what proxies represent or how they move. Camera, trajectories, timing, and object count still come from the source video.

Here is an original, untested Scene Intent for a two-shot underground product reveal:

The low rectangular proxy is a matte-black electric cargo bike with two amber headlamps. The tall moving proxy is its rider in a rust-colored rain jacket. The circular wall forms are wet concrete tunnel vents. Keep the scene grounded and realistic, with warm maintenance lights reflecting on the damp floor.

Notice what is missing. It does not ask for a dolly move, change the number of subjects, or introduce a third shot. Those decisions belong in the preview. The text only names unclear objects and establishes a coherent finish.

If the blockout already makes a relationship obvious, leave it out of the prompt. Repeating every visible movement creates more words without adding control.

Use references to lock appearance, not timing

Reference images are optional. The current fal product page allows up to eight and suggests using them for subjects, environments, interiors, or details. When supplied, those images are used directly as appearance guidance. Camera and movement still come from the video.

Choose references with a specific job:

• one clear view of the cargo bike's proportions and finish;
• one rider reference with visible jacket shape and color;
• one tunnel reference that establishes concrete, moisture, and light temperature.

Three compatible references are often easier to interpret than eight images that disagree about the same subject. A dramatic front view, a catalog side view, and a stylized illustration may each be strong on their own while pulling the generated object in different directions.

If no references are provided, the endpoint can generate new appearance references. The API field max_generated_reference_images defaults to two and allows a limit from one to eight. It may use fewer when it has enough coverage; shots without a generated reference remain in the video. Automatically generated references add to the bill, so supplying your own consistent images can improve both control and cost predictability.

Shot count can cost more than timeline length suggests

The detailed pricing block on fal's model page, checked October 17, 2026, lists a $0.50 processing fee per request, plus video generation at $0.05 per billed second for 480P, $0.08 for 768P, or $0.16 for 1080P. Each shot is rounded up to a whole second and has a minimum of five billed seconds.

That last sentence is the one to plan around.

A single five-second 768P shot starts at:

$0.50 processing + (5 × $0.08 video) = $0.90

A nine-second edit made from three three-second shots can be billed as 15 video seconds because every shot reaches the five-second minimum:

$0.50 processing + (3 shots × 5 billed seconds × $0.08) = $1.70

Both calculations exclude input-token overages and automatically generated appearance references. The detailed block includes 4,096 input tokens per shot, then lists $0.02 per 1,000 additional tokens; it lists generated references at $0.10 each.

fal currently displays conflicting helper copy elsewhere on the same page: it calls a five-second 768P job with supplied references “about $0.96” and estimates automatically generated references at $0.02–$0.04 each. The detailed block instead prints a $0.90 base example and $0.10 per generated reference. Treat the detailed block as a planning reference, inspect the live quote before submitting, and reconcile the final invoice rather than assuming either published figure is the transaction price.

The practical lesson is not “never cut.” It is to make each cut worth paying for. A nine-second one-shot camera move and a nine-second six-cut montage are different production and billing problems.

A useful first pass is intentionally boring

Do not test identity, intricate motion, six references, fast cuts, and final resolution in one expensive request. Prove the control chain in layers.

Start with one subject, one environment, and one continuous camera move. Export a readable blockout. Add one or two appearance references. Write a Scene Intent that only names the proxies. Render the first pass at 480P or 768P.

Then review it in this order:

1. Did the camera direction and framing survive?
2. Did each proxy become the intended object or character?
3. Did the timing and edit remain intelligible?
4. Did the references influence the right parts of the scene?
5. Is the remaining error structural, semantic, or visual?

Fix a structural error in the blockout. Fix a semantic error in Scene Intent. Fix an appearance error with a clearer reference. Changing all three inputs at once hides which decision actually mattered.

The source is strong guidance, not a geometry lock. fal explicitly notes that exact geometry, camera paths, motion, and reference details are not guaranteed. If a wheel must contact one precise point, a hand must grasp a product at an exact frame, or a logo must stay dimensionally perfect, treat the result as a concept or previs finish—not as a guaranteed final VFX composite.

When standard H3 Max is the better starting point

3D to Video is valuable when the shot is already directed in 3D. If all you have is an idea, building a full blockout may be unnecessary.

Use H3 Max image to video when one strong first frame can establish the composition and the main task is inventing motion. Use the storyboard-to-video workflow when you have static panels rather than an animated previs. Use the camera movement prompt guide when you want to direct a generated shot in words rather than inherit a camera path from previs. Use 3D to Video when blocking, camera, motion, and cuts are decisions you have already made and want the generated finish to follow as generative guidance.

This site's generator does not expose minimax/h3-max/3d-to-video and cannot accept a 3D preview video. It can create a standard H3 Max Turbo text-to-video or image-to-video clip at 480P or 768P while you explore the look of a scene before committing to a blockout.

The central rule is simple: put spatial and timing guidance in the video, identity and atmosphere clarifications in Scene Intent, and appearance guidance in references. When each input has one job, the workflow becomes easier to direct, easier to diagnose, and much easier to budget.