Reference image
Visible surfaces establish the subject, materials and scene layout.
H3 Max Camera Controls
Turn one image into a camera-controlled video.
Official fal Playground example · 768p · 10 keyframes
Source: falSet the orbit, height and distance of the camera. Keep the reference scene frozen while the viewpoint moves.
Explore camera pathsWhat is H3 Max Camera Controls?
Also called H3 Max Multi Angle, this endpoint generates a video from one reference image and a sequence of camera poses. Azimuth moves around the subject, elevation changes the viewing height, and distance moves the camera closer or farther away. The public response is a video file; it does not expose a mesh, point cloud or Gaussian splat.
Camera keyframes
Azimuth, elevation and distance specify the viewpoint at normalized times.
Generated viewpoints
The model synthesizes unseen views into a video. Hidden areas are inferred, not measured.
Design the camera movement
Choose a path, select a keyframe and adjust the camera. This diagram previews the path locally; it does not generate new H3 Max frames or call a paid API.
360° orbitProposed settings
0°Azimuth0°Elevation1.00×Distance
Edit keyframe
A complete turn around a fixed subject. Keep the signed 0° → 360° path to preserve a full revolution.
The table shows seconds. The API uses normalized time from 0 to 1; exported JSON converts it automatically. Distance is relative scene scale, not meters.
| Time | Azimuth | Elevation | Distance |
|---|---|---|---|
| 0.00 s | 0° | 0° | 1.00× |
| 3.00 s | 180° | 0° | 1.00× |
| 6.00 s | 360° | 0° | 1.00× |
Request JSON
{
"image_url": "https://pub-1ab9d7cc8a464826beace72316d08b27.r2.dev/h3-camera/v1/product.jpg",
"prompt": "Keep the single sneaker and its stone plinth rigid and motionless. Preserve the laces, sole layers, woven upper and green panels. Follow only the specified camera path. Keep the lighting fixed and add no objects or writing.",
"prompt_expansion_mode": "balanced",
"resolution": "480P",
"duration": 6,
"seed": 4201,
"camera_trajectory": [
{
"time": 0,
"azimuth": 0,
"elevation": 0,
"distance": 1
},
{
"time": 0.5,
"azimuth": 180,
"elevation": 0,
"distance": 1
},
{
"time": 1,
"azimuth": 360,
"elevation": 0,
"distance": 1
}
],
"enable_safety_checker": true
}Six scenes to test
Six original input images and reproducible camera recipes. Open a scene for its prompt, settings and inspection checklist. A recipe is labelled as unrun until a real H3 Max output is attached.
Freeze a leap; move the camera around it.
Not yet generatedTurn a product hero image into multiple viewing angles.
Not yet generatedInspect facial identity and garment continuity.
Not yet generatedTest spatial consistency around furniture.
Not yet generatedReveal the side and back of a parked car.
Not yet generatedChange viewing height around a suspended trick.
Not yet generated
Requested duration: 6 seconds. The supplied media lasts 6.58 seconds. fal reports 4.84 seconds of model inference for this example; this excludes the full queue, prompt expansion and delivery time. Source: fal
Camera path recipes
Official launch & more H3 Max videos
The fal launch concerns Camera Controls. The other links discuss H3 Max more broadly; they are not evidence of this endpoint's camera accuracy. View counts and generation speeds vary over time.
Is it a world model?
H3 Max Camera Controls is useful for spatially aware video rendering. Its public API delivers video, so treat it as a camera-controlled video generator rather than an exported 3D world.
Camera Controls / Normal image-to-video
Standard image-to-video usually describes motion in a text prompt. This endpoint adds explicit camera keyframes and a frozen-scene default. It offers a more concrete path specification, while subject and geometry consistency still need visual inspection.
| Capability | H3 Max Camera Controls | World Labs Atlas |
|---|---|---|
| Reference inputs | Single image | One or more images; other modalities in the research announcement |
| Camera control | Azimuth, elevation, distance keyframes | Native camera geometry |
| Delivered output | Video URL, expanded prompt and timings | Research demos show images, video and explicit 3D |
| Point clouds / splats | Not exposed by this endpoint | Explicitly demonstrated |
| Simulation | No simulation interface documented | Space-time and robotics simulation demonstrated |
| Access | Public paid fal endpoint | Early access; separate from the public Marble World API |
Atlas is in early access. World Labs already has a public World API for Marble; that does not make Atlas a generally available API.
Parameters that matter
Input
1 imageReference image plus an optional motion prompt.
Camera path
2–12 keyframesElevation −90° to +90°; positive relative distance.
Duration
5–15 secondsInteger seconds in the current API schema.
Resolution
480p · 768p · 1080p1080p uses latent refinement from native 768p.
Prompt expansion
balanced / qualityA richer rewrite can add preparation time.
Usage
Commercial use listedGeneration is paid through fal; terms apply.
Source: fal API · Last checked: 2026-09-12
Pricing & cost calculator
Prices recorded on September 12, 2026. fal lists a 75% launch discount ending September 14. The regular prices below are its announced post-promotion rates; check fal before running a job.
| Resolution | Recorded launch price | Announced regular price |
|---|---|---|
| 480p | $0.0125 / s | $0.05 / s |
| 768p | $0.02 / s | $0.08 / s |
| 1080p | $0.04 / s | $0.16 / s |
Last checked: 2026-09-12 · Source: fal
Estimates use output duration, resolution and clip count. They exclude retries and separate image-generation costs. This calculator does not charge your account.
What to inspect in the result
Hidden surfaces
A single image cannot establish the true back side of an object. New views can contain invented details.
Geometry drift
Watch for changing window counts, bent furniture, irregular wheels and shifting proportions.
People and frozen motion
Check faces, hands, floating objects and loose clothing for unwanted movement.
Complex scenes
Thin structures, reflections and overlapping foreground objects can break continuity. Begin with small camera moves.
Use the camera path in your API request
Export a path above, set image_url to your reference image and submit the JSON to the fal queue. Keep FAL_KEY on your server. Use the returned status_url and response_url to track and retrieve the job.
minimax/h3-max/camera-controls
curl --request POST \
--url https://queue.fal.run/minimax/h3-max/camera-controls \
--header "Authorization: Key $FAL_KEY" \
--header "Content-Type: application/json" \
--data @h3-camera-orbit.jsonEach keyframe needs time, azimuth, elevation and distance. time is between 0 and 1, elevation is between −90 and +90, and distance must be greater than zero.
fal APIFrequently asked questions
Is H3 Max Camera Controls free?
The fal endpoint is paid. You can watch this page's example and edit camera paths without an API call. For free public model demos, open our MiniMax H3 page.
Is Multi Angle the same model?
The fal Camera Controls page also uses the name H3 Max Multi Angle. This guide refers to the minimax/h3-max/camera-controls endpoint.
Can I export a real 3D model?
This endpoint returns a video, an optional expanded prompt and timing data. Its public output schema does not include a mesh, point cloud or Gaussian splat.
Does 4.84-second inference mean every job finishes that fast?
No. That value belongs to one official example and measures backend inference. Queueing, prompt expansion and delivery can add time.
Can I make a full 360° product video?
The schema preserves signed full azimuth turns. A complete orbit can be specified, but unseen surfaces are generated guesses. Check product geometry before commercial use.