RealMe: Immersive Volumetric Video from Minimal Captures
RealMe turns sparse, low-resolution captures into 2K video designed to stay consistent across viewpoints and time.
The framework adapts a high-resolution video diffusion model to the capture cameras’ layout, then uses the enhanced videos to refine a volumetric reconstruction. A frequency-based training loss helps narrow the gap between generated footage and real captures. The authors report better visual fidelity and consistency across views and time than existing methods on the public DNA-Rendering benchmark and their own RealVV dataset.
HF Daily Papers' note
The framework adapts a high-resolution video diffusion model to the capture cameras’ layout, then uses the enhanced videos to refine a volumetric reconstruction. A frequency-based training loss helps narrow the gap between generated footage and real captures. The authors report better visual fidelity and consistency across views and time than existing methods on the public DNA-Rendering benchmark and their own RealVV dataset.
HF Daily Papers' note
score 5