Consistent 3D Object Insertion in Image-to-Video Generation via Render-Conditioned Diffusion

Producing photorealistic advertising footage from digitally scanned 3D products requires placing an object into a virtual background scene and preserving its geometric and photometric identity as the camera moves. While acquiring high-fidelity 3D product scans has become increasingly accessible, automatically integrating them into video backgrounds at scale remains an open problem. Current image-to-video diffusion models lack an explicit geometric anchor: without one, the inserted object progressively degrades and drifts as the viewpoint changes, making it unsuitable for advertising applications. We present a training-free, inference-time pipeline that takes a single background image and a physically-based rendering (PBR) 3D mesh asset as input and produces a geometrically consistent video in which the inserted object maintains its appearance throughout the camera trajectory. A background mesh reconstructed from monocular depth estimation replaces the sparse point-cloud representation of baseline approaches, providing the topological connectivity required for correct shadow casting and occlusion. Path tracing under HDRI illumination estimated from the background image produces contact shadows and physically accurate material responses that visually ground the object in the scene. To prevent temporal drift over sequences of arbitrary length, an autoregressive block-based anchoring step recalibrates the scene depth at every N-frame boundary using the object's exact rendered depth map as a metric constraint. Evaluated over 6 benchmark background scenarios, the Object-Driven Autoregressive (OD-AR) pipeline significantly outperforms both a point-cloud baseline (640x reduction in object reconstruction error) and a non-autoregressive mesh baseline (28x improvement in object reconstruction error and improved object placement accuracy), while bounding temporal drift to a stable saw-tooth profile rather than allowing it to diverge monotonically.