Cloud Training Compared: Luma, Zhitianxia and Pointcosm

Luma AI offers the lowest-barrier cloud training entry: upload video or image sequences via mobile app or web browser with zero local configuration. After upload, the job enters a queue and typically completes in 10-30 minutes. On failure, the platform gives a brief explanation — usually poor image quality or incomplete upload — and you can resubmit directly. Zhitianxia targets more professional B2B users with batch upload and project management, but the registration and billing flow is more complex, suited to teams with steady throughput needs. Pointcosm sits between the two, supports image and video dual-mode, with a Chinese-language interface that's friendly to domestic users. [unverified]

The three platforms use different pricing models: Luma is primarily subscription-based, with free quotas for initial trials and paid plans for professional use; Zhitianxia bills mainly per-project or per-volume, suited to commercial projects requiring predictable per-job costs; Pointcosm offers prepaid packages by image count or duration (publicly cited tiers include 300/900 images and 3/9 minutes — verify current pricing on their site). Beyond unit price, also factor in queue times, failure-retry costs, and export format support before deciding. [unverified]

Export format support varies by platform — confirm your downstream workflow's requirements before choosing. PLY is the most universal format, compatible with SuperSplat, Postshot and most tools, lossless but bulky; splat is a compact binary format roughly half the size of PLY; spz is Niantic's flagship compressed format, even smaller but requiring a dedicated decoder. For web display, confirm your renderer supports the target format first; for archival or retraining, keep a PLY master. Some platforms export only proprietary formats — verify before committing.

Small objects or single rooms (under 100 images): Luma's free quota is usually sufficient and offers the fastest onboarding. Medium scenes (100-500 images, full interior or small outdoor): all three platforms can handle this — run a small batch test before committing at scale. Large scenes (500+ images, building exterior or campus): prioritize platforms with a professional or enterprise tier, confirm per-job image limits and priority queue access. If the scene involves data security or privacy requirements, cloud paths need careful evaluation — local training may be the more appropriate choice.