Depth Estimation

The depth estimation stage generates a per-pixel depth map for every frame. The depth map encodes how far each pixel is from the virtual camera, enabling stereo synthesis downstream.

Model selection

anelo supports three depth model architectures. Each has different strengths:

Depth Anything V2 is the default and recommended model. It produces high-quality depth maps across a wide range of content types — natural scenes, interiors, close-ups, and wide shots. It handles occlusion and complex geometry well.

MiDaS is a mature model with broad compatibility. It runs faster than Depth Anything V2 but produces slightly less detailed depth on complex scenes. Good for batch processing where speed matters more than maximum depth precision.

ZoeDepth produces metrically accurate depth maps — the absolute distances between objects are more accurate, not just the relative ordering. Useful for content where metric accuracy matters (architectural visualization, spatial computing).

Temporal smoothing

Video depth estimation must be temporally consistent — adjacent frames should have similar depth maps unless there is real scene change. The temporal smoothing parameter controls how much inter-frame consistency is enforced. A value of 0 disables smoothing (each frame is independent). A value of 1.0 applies maximum smoothing. The default of 0.5 provides good consistency without blurring fast motion.

Edge refinement

Depth maps can have soft boundaries between foreground and background objects. Edge refinement sharpens these transitions, producing cleaner stereo output at object boundaries. The trade-off is slightly longer processing time per frame.