QWE AI Academy Type “how to use Luma AI for 3D video” into a search bar and you get two completely different answers shoved into the same article. One is about scanning a real object with your phone and getting a spinnable 3D model. The other is about typing a prompt and getting a generated video clip with believable depth and parallax. These are not the same product. They live in different apps, cost different credits, and solve different problems.
If your goal is a real 3D asset you can rotate, embed, or hand to a 3D artist – you want Luma 3D Capture (NeRF / Gaussian Splatting). If your goal is a short cinematic clip with motion that looks 3D – you want Dream Machine. Picking the wrong one wastes hours. Most tutorials skip this fork entirely, which is why people end up frustrated.
The reader scenario: which “3D video” do you actually need?
Here’s the quick test. Will someone need to interact with the result – spin it, walk through it, drop it into Blender or a game engine? Then you need a 3D capture, exported as a flythrough video or as a model file. Will the final deliverable be a flat MP4 that plays once and ends? Then Dream Machine is faster, cheaper per second, and won’t make you walk in slow circles around a chair.
| You want… | Use this | Output |
|---|---|---|
| A spinnable 3D scan of a real object/space | Luma iOS/Android app or web upload (3D Capture) | NeRF, Gaussian Splat, .obj/.gltf/.ply |
| A flythrough video of that scan | 3D Capture → render a camera path | MP4 (flat video of a 3D scene) |
| An AI-generated cinematic clip | Dream Machine (Ray3 model) | 5-10 sec MP4 |
Most published tutorials only show the third row. That’s fine if you want stock footage. It’s useless if a client asked for a turntable of their product.
Path A: Scanning a real object with Luma 3D Capture
This is the older, more technical Luma. Luma Labs popularized NeRFs (Neural Radiance Fields), then moved most users to Gaussian Splatting because splats render in real time at roughly 60fps in a browser, while NeRFs are heavy to view. The capture process is the same either way – your phone’s AR tracking does the spatial math while you walk.
The Loop Method is the one technique you can’t skip. Three passes around the subject:
- Eye level – overall shape and silhouette.
- Low angle, looking up – captures undersides, edges, anything occluded from above.
- High angle, looking down – captures the top and any horizontal surfaces.
Move slowly. The phone uses AR tracking to know where it is in space, and if you swing it too fast the tracking slips silently – you don’t get an error, you get a warped model with the back of the chair fused into the wall. The fix is always the same: re-shoot, half speed.
Pro tip: Don’t scan in changing light. A cloud passing over a sunlit object mid-loop is one of the most common reasons captures look “melted.” Either go full shade, full overcast, or indoor with steady lights. Reflective surfaces (chrome, glass) are fine – that’s actually where NeRFs beat traditional photogrammetry – but inconsistent lighting is fatal.
Once the capture finishes processing, Luma gives you a camera-path editor. You set keyframes for a virtual camera, hit render, and out comes an MP4. That is your “3D video” – a flat video file showing a real 3D scene. You can also export the underlying scene as .obj, .gltf, or .ply if a 3D artist is going to take it from there.
Here’s a question worth sitting with: why does the 3D Capture path stay almost invisible in mainstream AI coverage, even though it predates Dream Machine and produces genuinely different outputs? Probably because scans require a real-world subject, patience, and a phone – none of which make for an exciting Twitter demo. Generated video is instant and shareable. The boring tool is often the more precise one.
Path B: Generated 3D-feeling video with Dream Machine
Dream Machine is the product most people now mean when they say “Luma AI.” Per Luma’s product page, it’s a video model trained on motion and physics rather than just frames – which is why pans, parallax, and camera moves feel volumetric instead of like a warping image. It launched publicly in June 2024 and has crossed 25M registered users (per Luma’s own marketing – verify if exact figures matter for your reporting).
Setup is short. Sign in with Google at lumalabs.ai, pick “Dream Machine,” and you land on a prompt box. Two modes:
- Text-to-video – describe the scene from scratch.
- Image-to-video – upload a start frame, optionally an end frame, and prompt the motion between them.
Image-to-video is what you want for anything that needs to look specific. Generate (or shoot) the exact frame you want in Photon, Midjourney, or a real camera. Upload it. Then prompt only the motion: “slow dolly-in, camera arcs left, light shifts warmer.” Leaving the prompt blank usually gives you a generic camera pan – the model needs you to tell it what’s moving.
The End Frame feature takes this further. You give it frame A and frame Z, and the model interpolates the middle. It’s the cleanest way to get a controlled transition without a roll of the dice on each generation.
The credit math nobody spells out
Luma’s pricing looks simple – Free, Lite, Plus, and up – but the credit costs per generation vary wildly by model and resolution. Here are the actual numbers from Luma’s credit documentation (current as of early 2026 – check the page before budgeting a project):
| Generation | Credits |
|---|---|
| Photon image (per image, batches of 4) | 4 (16/batch) |
| Ray3 540p SDR, 5s | 160 |
| Ray3 540p SDR, 10s | 320 |
| Ray3 HDR, 10s | 1,280 |
| Ray3 HDR + EXR, 10s | 2,240 |
Read that bottom row carefully. A single HDR+EXR 10-second clip costs 2,240 credits. If you’re delivering professional HDR work, check Luma’s current plan page before assuming any mid-tier subscription covers a full project – the numbers may surprise you. For that kind of output, you likely need Unlimited or top-up credits.
Speaking of top-ups: monthly credits do not roll over, but Top-Up Credits (starting at $4 for 1,200 credits, as of early 2026) stay valid for 12 months and even survive a downgrade to Free. For sporadic users this is genuinely better than upgrading a tier.
The gotchas tutorials don’t mention
Web credits ≠ API credits. Subscribe to Plus on the web, then try the Dream Machine API – your credits won’t be there. Zero. Luma’s docs state this plainly, but it’s still the thing developers hit first and email support about second. API billing is a completely separate purchase; the subscription doesn’t touch it.
Lite plan is non-commercial (as of early 2026, per Luma’s payment docs). The $9.99/month tier produces watermarked output restricted to personal use. Drop a Lite-plan clip into a client deck and you’re in license-violation territory – cropping the watermark doesn’t change the terms. The commercial entry point is Plus at $29.99/month.
Text in scenes will look like alien runes. Like every current AI video model, Dream Machine cannot reliably render readable text. The community workaround: build the sign or label in Photoshop, set it as your start frame in image-to-video, and prompt only the camera move – never ask the model to change the text.
What about combining the two?
This is the actual interesting question. You can scan a static object with 3D Capture and then animate something around it in Dream Machine using image-to-video – but there’s no native pipeline that takes a Luma splat and makes it walk. The convergence everyone keeps writing about is still mostly aspirational. Worth watching, not worth planning a project around yet.
FAQ
Can I get a real 3D model file out of Dream Machine?
No. Dream Machine outputs flat video. Only the 3D Capture side of Luma exports model files (.obj, .gltf, .ply).
Why does my Dream Machine clip look great for 4 seconds and then fall apart?
Most likely you used the Extend feature. Each extension takes the last frame as the new starting frame, so small drifts compound – the character’s face slowly stops being the same face. Cap chained extensions at two, and if you need a longer continuous shot, generate fresh image-to-video clips and edit them together rather than extending. For one-off generations, the 5-second SDR option is also more stable per-credit than 10-second.
Is the Free plan enough to learn on?
For 3D Capture, mostly yes – basic scans work. For Dream Machine, the Free tier limits you to draft-resolution video and watermarked images, so you’ll outgrow it the first day you try anything serious.
Next step: open the Luma app on your phone right now and scan one object on your desk using the Loop Method. Don’t read another tutorial first. The 3D Capture workflow is the half of Luma that almost nobody teaches well, and ten minutes of a real scan teaches more than an hour of articles.