An advanced, real-time stereo rendering engine for 2D-to-3D conversion in VR and stereoscopic displays.
VisionDepth3D introduces multiple first-of-their-kind techniques in the 3D conversion community, while also integrating established practices into a single GPU-optimized pipeline.

• The shape_depth_for_pop algorithm is unique to VisionDepth3D:
  1. Percentile stretch of depth range.
  2. Recentered on subject depth.
  3. Symmetric gamma curve for controlled “pop”.
• Enables tunable cinematic or VR-style stereo with consistent subject emphasis.

• Dynamically estimates subject depth (center-weighted histograms + percentiles).
• Locks the zero-parallax plane to the subject with exponential smoothing.
• Prevents subject “drift” and stabilizes the screen plane without manual keyframing.

• Computes normalized variance of depth in the central region:

parallax_scale = compute_dynamic_parallax_scale(depth_tensor)

• Adapts stereo strength automatically — gentle for flat shots, expansive for landscapes.

• Novel use of depth gradients to suppress parallax shifts near thin, detailed edges:

edge_mask = torch.sigmoid((grad_mag - edge_threshold) * feather_strength * 5)

• Prevents ghosting and halos without AI inpainting.

• Combines distance transforms with EMA smoothing to “round” depth on subjects.
• Prevents shimmer in hair, fingers, and soft edges.
• Creates natural curvature without requiring heavy segmentation pipelines.

• Subject-aware floating window automatically detects window violations.
• FloatingWindowTracker + FloatingBarEaser smooth jitter and clamp drift.
• First real-time system that applies cinematic floating windows automatically.

• Depth-of-field blur controlled by a focal depth tracker that adapts to scene motion.
• Busy shots → faster focus shifts; still shots → stable focus.
• Simulates real cinematography rules dynamically during 3D rendering.

• Detects warping gaps via gradient magnitude and fills with blended original + blurred content.
• Lightweight alternative to neural inpainting — seamless and invisible.

• Depth-weighted continuous parallax shifting – smooth stereo gradients instead of discrete layers.
• GPU tensor grid warping – CUDA-optimized grid_sample per-eye rendering.
• Scene-aware dampening – adjusts disparity for flat vs. complex scenes.
• Temporal percentile EMA normalization – stabilizes depth scale across frames.
• Depth-based DOF (multi-level Gaussian pyramid) – established technique, enhanced in VD3D with motion-aware focus.
• Black bar detection & aspect handling – auto-crop and cinematic aspect preservation.
• Color grading & sharpening – GPU-accelerated saturation, contrast, brightness, and safe sharpening.
• Multi-format 3D output – Half-SBS, Full-SBS, VR, anaglyph, interlaced.
• FFmpeg streaming codec pipeline – CPU (libx264/x265/AV1), NVIDIA NVENC, AMD AMF, Intel QSV with CRF/CQ control.

VisionDepth3D combines these original contributions with proven practices to form a holistic, real-time, GPU-accelerated 2D→3D pipeline.

📄 Licensed under: VisionDepth3D Custom Use License (No Derivatives)
🔗 Project: https://github.com/VisionDepth/VisionDepth3D