Skip to content

Changelog.md

Latest commit

 

History

History
145 lines (95 loc) · 5.77 KB

Changelog.md

File metadata and controls

145 lines (95 loc) · 5.77 KB

VisionDepth3D v3.9 - Changelog

This release introduces major upgrades to VisionDepth3D’s model loading, depth workflow, VR180 output, and pipeline stability, with improved Hugging Face integration, ONNX support, and overall reliability across the app.

3D Video Generator

VR180 Equirect Output (Initial Release)

VisionDepth3D now supports native VR180 equirect stereo output.

New Output Modes

  • Added VR180 Equirect (Top-Bottom)
  • Added VR180 Equirect (Side-by-Side)

Dual-Resolution Workflow

Introduced a two-stage VR180 resolution system:

  • Flat Working Resolution (Per Eye) – Internal render resolution used for stereo warping and depth-based effects.
  • Equirect Per-Eye Resolution (2:1) – Final projected resolution used for VR headset playback.

This allows users to independently balance performance and final visual sharpness.

VR180 Controls

  • Added adjustable HFOV (60°–140°)
  • Added manual Equirect WxH configuration
  • Added manual Flat Working WxH configuration
  • Preset-ready structure for common VR180 resolutions

SDR & HDR Support

  • VR180 output works in both SDR and HDR10 rendering pipelines
  • HDR10 VR180 maintains correct 10-bit encoding and color metadata through FFmpeg

Projection & Stereo Packing

  • Proper hemispherical projection per eye
  • Correct stereo packing for both TB and SBS modes
  • Consistent output sizing across FFmpeg and OpenCV rendering paths

VR Player Naming Compatibility

  • Updated auto-naming for VR180 outputs to align with VR players:
    • _SBS_180
    • _TB_180

Ensures correct automatic detection in players such as DeoVR.

FPS/Upscale Enhancement (formerly FrameTools)

Renamed & Reworked

  • Renamed FrameTools to FPS/Upscale Enhancement for clearer purpose and usability

UI Improvements

  • Redesigned the FPS/Upscale Enhancement tab layout for a cleaner, more intuitive workflow
  • Improved spacing, grouping, and visual hierarchy for faster setup and easier readability
  • Introduced Pause, Resume, and Stop buttons

Performance & Stability Improvements

  • Fixed the threaded RIFE + ESRGAN processing pipeline that previously failed to run correctly
  • Matched threaded pipeline behavior to the merged pipeline (native-resolution RIFE with final-stage upscaling)
  • Removed unnecessary resizing and duplicate processing that caused slowdowns and memory spikes
  • Improved frame queue handling for smoother throughput and better GPU utilization
  • Fixed progress reporting and ETA calculation to prevent incorrect values and UI update spam
  • Added safe, bounded frame loading with cancellation support to prevent deadlocks and excessive RAM usage
  • Stopped swallowing queue exceptions by handling Full and Empty normally and surfacing unexpected errors for easier debugging
  • Prevented the writer reordering buffer from growing indefinitely if a frame ID goes missing by adding a skip/limit safeguard
  • Improved FFmpeg writer robustness by validating frame format/size before writing, reducing crashy edge cases

Additional Pipeline Improvements (Post-Fix Enhancements)

  • Implemented dynamic model loading system using Hugging Face repositories
  • Removed requirement to ship large model weights with the application
  • Models are now downloaded on demand and cached to the local weights/ folder
  • Ensured all runtime model loading resolves correctly regardless of install method
  • Refactored model initialization to prevent duplicate loading and improve startup reliability
  • Removed redundant frame normalization passes that caused unnecessary CPU overhead
  • Simplified threaded pipeline frame flow to reduce queue overhead and unnecessary data copying

Performance Insight

  • Identified input resolution as a primary performance bottleneck during RIFE + ESRGAN processing
  • Lowering input resolution significantly improves processing speed due to reduced pixel workload
  • Input Resolution % setting now serves as a key control for balancing performance vs quality

Speed Improvements

  • Additional optimizations reduce CPU overhead and improve GPU utilization during processing

Preview GUI

Stability Fix

  • Fixed issue where the Preview GUI window would not properly close in .exe builds
  • Ensured proper window cleanup and destruction to prevent hanging UI processes

Depth Engine

  • Fixed packaging for DA3 in the .exe build
  • Fixed Process Video Folder in the depth estimation pipeline
  • Fixed UI freezing during depth folder processing and restored live progress bar updates
  • Refactored depth folder and single-video processing to safely separate UI-side controls from background worker execution
  • Fixed depth pipeline argument handling for single video and folder-based processing
  • Added support for loading ONNX depth models from either local model folders or Hugging Face repositories
  • Restored ONNX-specific warm-up and inference handling for Hugging Face-hosted ONNX models
  • Hid diffusion-only controls such as Inference Steps and CPU Offload Mode unless a compatible diffusion model is selected
  • Improved Hugging Face ONNX model detection for Distill-Any-Depth and Video Depth Anything exports
  • Cleans up Letterbox.json file after rendering so no random temp file is left behind

Special Thanks

A big thank you to AcolyteOfHedone for contributing fixes and technical improvements that helped strengthen this release, including AMD AMF encoder fixes, ONNX adjustments, and AMD GPU provider compatibility work.

GitHub: EvolvingProficiency

Upgrade Note
Back up your weights/ and presets/ folders before uninstalling v3.8.2 Then run VisionDepth3D_Setup_Downloader to download the official
VisionDepth3D v3.9 Windows installer and required .bin files.