VisionDepth3D v4.2.2 Changelog Depth Anything V2 Performance Fix
Fixed a major DA-V2 performance issue when using the Original inference resolution.
DA-V2 no longer treats Original as full-resolution model inference by default.
Original now preserves the source aspect ratio while capping the DA-V2 model input size to a safer default.This prevents accidental 1080p/4K DA-V2 inference, which could severely reduce render speed.
Added environment override for users who intentionally want true original-resolution DA-V2 inference:
VD3D_DAV2_DEFAULT_MAX_SIDE=0
Default DA-V2 max side is now:
VD3D_DAV2_DEFAULT_MAX_SIDE=518 DA-V2 Adapter Optimization
Reduced unnecessary DA-V2 output resizing inside the adapter.
DA-V2 now returns model-resolution depth output and lets the main VisionDepth3D postprocess handle final video/image resizing.
Removed duplicate DA-V2 normalization before main pipeline normalization.
Reduced GPU work and GPU-to-CPU transfer size during DA-V2 video rendering.
Improved DA-V2 handling for Original resolution while keeping aspect ratio.
Added clearer DA-V2 load diagnostics showing:
selected weight file
backend
device
dtype
FP16 state
Added main load-setting diagnostics showing:
selected model
checkpoint
UI FP16 state
resolved FP16 state
active backend
torch device
CUDA FP16 still requires enabling FP16 before loading/reloading the model.
DA-V2 Original mode is now optimized for practical render speed by default.
True original-resolution DA-V2 inference remains available through the environment override, but may be very slow on high-resolution video.