updream's Previs Studio brings 3D white-model previsualization to AI video creation

When a 2,000-character prompt still cannot control the camera, a new answer is emerging: build a 3D white model first. Chinese AI video tool updream has launched Previs Studio, a feature that moves the previsualization workflow of the film industry into its creation canvas, according to a hands-on report from QbitAI.

The workflow is "design the shot first, generate the picture second": creators build a 3D white-model scene, set camera positions, plan character movement, record a previs video, and hand it to video models such as Seedance, Kling, Wan, or Gemini Veo for final rendering — solving the "what to shoot" and "how to shoot" control problems of AI video.

In practice, users upload one to three scene reference images and the AI generates a high-fidelity 3D white-model scene in minutes (the official estimate is four to seven minutes), with 360-degree rotation and navigation. Creators can then add characters and cameras — character motion supports global and local actions — draw movement paths and tracks, and record the previs for export to downstream generation.

The white model carries pure spatial and temporal information. It acts as a spatial anchor that helps the model understand depth, occlusion, proportion, and movable space within a shot — concepts that text prompts describe poorly. updream defines white-model control as an advanced capability and ships dedicated prompt specifications for coarse-grained and fine-grained usage.

QbitAI's test covered three notoriously hard shot types: one-take sequences, multi-camera cuts, and complex camera moves. With white-model reference, character direction, camera paths, and final resting positions were fixed far more reliably; multi-camera scenes kept character identity, spatial orientation, and environment continuity stable, cutting wasted attempts.

The idea itself is not new — ControlNet in the Stable Diffusion era used depth maps and normal maps to constrain image generation, and the approach has migrated to video. What changed is accessibility: building a white model once required months of Blender or Maya training; updream's AI collapses that barrier.

Previs Studio is not magic: white-model fidelity depends on reference image quality, complex motion still needs patient fine-tuning, and final renders can still suffer broken character consistency or physical detail glitches that need local re-editing. But the deeper shift is in creative control: creators stop guessing how the model will shoot and decide how they want to shoot first.

In the industry context, after half a year of intense competition among Chinese video models, basic generation capability is no longer the differentiator — the next battleground is stable control over results. updream is pairing the launch with a month-long Previs Studio Challenge offering up to tens of thousands of yuan in rewards across showcase and tutorial tracks. The open question is whether white-model previsualization becomes a standard AI video workflow, and whether other platforms follow.