Scene Light AI Information

Scene Light is a lighting tool that leverages the Depth Anything V2 and Video Depth Anything AI models to generate depth data, and the Microsoft MoGe-2 model to generate normals for footage and images. The models are self-contained files installed locally on your machine. They do not learn from or share any content processed through them, nor do they rely on external servers or cloud-based processing. This means no data processed by the models leaves your machine.

How the models were trained

Scene Light uses three locally installed AI models, each trained with a different approach suited to its task:

1. Depth Anything V2: Depth Anything V2 was trained for single-image depth estimation using a combination of synthetic images with exact depth labels and real-world images that did not originally include depth information. To make the real images useful for training, the team used a teacher model to estimate their depth and create pseudo-labels. The final student model was trained on this mix of precise synthetic data and pseudo-labeled real imagery, helping it generalize better to real-world scenes.
2. Video Depth Anything: Video Depth Anything builds on Depth Anything V2 and adapts it for video. The model replaces the image-only head with a spatial-temporal head designed to produce more consistent depth across frames. It was trained on a joint dataset of video depth data and unlabeled images, similar to Depth Anything V2, and uses a temporal consistency loss that constrains changes in depth over time. This helps reduce flicker and improves consistency across long video sequences without requiring optical flow, camera pose data, or other geometric priors.
3. The training material for Microsoft MoGe-2 was sourced from a massive, mixed corpus consisting of 24 individual datasets.

To build a robust model capable of open-domain geometry estimation, the researchers pulled data across three primary structural categories:

• 16 Synthetic Datasets: Used to provide mathematically flawless, razor-sharp geometric baselines.
• 3 LiDAR Datasets: Used to supply highly accurate real-world metric scales.
• 5 Structure-from-Motion (SfM) Datasets: Used to incorporate diverse, real-world multi-view geometry environments.

Together, these training approaches explain why the models behave differently: Depth Anything V2 is optimized for general single-image depth, Video Depth Anything extends that foundation for temporally consistent video depth, and Microsoft MoGe-2 is designed for open-domain geometry estimation using a broad mix of synthetic, LiDAR, and Structure-from-Motion datasets to support accurate normals with metric scale and sharper detail.