VisionDepth3D v3.9.0
VisionDepth3D v3.9 is a major update focused on workflow stability, model delivery, and overall usability across the 3D, depth, and enhancement pipelines.
Key Highlights
VR180 equirect output support
This release introduces native VR180 equirect stereo output with support for both Top-Bottom and Side-by-Side formats. A dual-resolution workflow was added so flat working resolution and final equirect output resolution can be controlled independently for better performance and quality tuning.
FPS/Upscale Enhancement improvements
The old FrameTools workflow has been renamed and reworked into FPS/Upscale Enhancement. This update includes UI cleanup, pause/resume/stop controls, and major fixes to the threaded RIFE + ESRGAN pipeline for better throughput, safer queue handling, cleaner progress reporting, and improved FFmpeg writer stability.
Hugging Face model delivery
VisionDepth3D now supports downloading and loading supported models directly from Hugging Face instead of relying as heavily on bundled model files. Models are cached locally after download, which reduces shipping size and improves long-term flexibility for updates.
Depth engine fixes and ONNX improvements
The depth pipeline received major fixes for single-video and folder-based processing, including UI freeze fixes, progress bar restoration, safer worker-thread handling, and improved FFmpeg output behavior. ONNX model loading now works from both local folders and Hugging Face repositories, with improved warm-up and compatibility handling for Video Depth Anything and Distill-Any-Depth exports.
Smarter UI behavior
Diffusion-only controls such as Inference Steps and CPU Offload Mode are now hidden unless a compatible diffusion model is selected, helping keep the depth UI cleaner and easier to understand.
3D pipeline reliability improvements
The 3D pipeline now includes better fallback handling for source video metadata when OpenCV fails to report valid FPS or dimensions, improving compatibility across more files and systems.
Notes
- Upscaling Models now downloaded from Hugging Face and cached locally after first download.
- ONNX and diffusion branches received targeted compatibility fixes in this release.
- This version is intended as a broader workflow update and not just a small patch release.
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
Full Changelog
For the complete update log, see the full changelog here: