An open-source alternative to Owl3D: convert any 2D video into stereoscopic 3D using AI monocular depth estimation (Depth Anything V2).
For each frame: estimate depth with Depth Anything V2 → keep the original frame as the left eye → warp the right eye by depth-based disparity → stream straight into ffmpeg. No temporary files, original fps and audio are preserved.
- AI depth estimation — Depth Anything V2 (small/base/large), GPU-accelerated (CUDA, Apple Silicon MPS); the
fastprofile runs on the Apple Neural Engine via Core ML on macOS (~3× faster than MPS) - Output formats — side-by-side (SBS), top-bottom, red-cyan anaglyph
- 3D TV mode — optional half-width SBS: each eye is squeezed horizontally for TV playback
- Stereo metadata — SBS and top-bottom outputs include H.264 frame-packing flags plus MP4
st3dor MKV StereoMode metadata for player detection - Apple spatial video — optional MV-HEVC export for Vision Pro (
--spatial) - Quality preservation — the left eye is the untouched source frame; audio is copied as-is
- Temporal smoothing — reduces depth flicker between frames
- CLI and GUI — command line or a simple Tkinter interface
Requirements: Python 3.10+, FFmpeg, and optionally a GPU (NVIDIA CUDA or Apple Silicon).
pip install -r requirements.txtThe depth model (~100 MB for fast) is downloaded automatically from Hugging Face on first run.
For Vision Pro spatial video output (macOS only):
brew install spatial# Basic conversion (side-by-side)
python main.py input.mp4
# Half-width SBS for a 3D TV (1920x1080 source -> 1920x1080 output)
python main.py input.mp4 --3dtv
# Choose output, format and depth model size
python main.py input.mp4 -o output.mp4 -f tb -p precision
# Stronger 3D effect, custom per-eye resolution
python main.py input.mp4 --strength 0.05 --eye-resolution 1920 1080
# Also produce an Apple spatial video for Vision Pro
python main.py input.mp4 --spatial| Option | Default | Description |
|---|---|---|
-p, --profile |
fast |
Depth model: fast (Small), balanced (Base), precision (Large) |
-f, --format |
sbs |
sbs, tb, or anaglyph |
--3dtv |
off | Half-width SBS for 3D TVs; SBS only, incompatible with --spatial |
--strength |
0.03 |
Max disparity as a fraction of frame width |
--smoothing |
0.3 |
Temporal depth smoothing (0 = off) |
--eye-resolution W H |
source | Resolution per eye before optional 3D TV squeezing |
--crf / --preset |
18 / medium |
x264 encoding quality/speed |
--spatial |
off | Also export MV-HEVC spatial video (SBS only, macOS) |
python main.py --guiSelect 3D TV mode (half-width SBS) with the SBS format. A 1920x1080
source produces two 960x1080 views in one 1920x1080 frame; select SBS 3D
playback on your TV to stretch each view back to full width. With a custom
--eye-resolution W H, the combined output is W x H in this mode, except
when W:H is narrower than 16:9: each eye receives centered black side bars
before squeezing, widening the combined output to 16:9 (rounded up to an
even width). For example, 1440x1080 becomes 1920x1080. Depth estimation and
stereo warping still process the original image without the bars. Images
at 16:9 or wider receive no padding.
Dimensions are rounded down to even values for encoding. Default SBS output
remains full width (3840x1080 for a 1920x1080 source). TV-mode filenames use
_stereo_hsbs.mp4.
from pipeline import Pipeline
Pipeline(profile="fast", strength=0.03).convert("input.mp4", "output.mp4", format="sbs")| Profile | Model | License |
|---|---|---|
fast |
Depth Anything V2 Small | Apache-2.0 |
balanced |
Depth Anything V2 Base | CC-BY-NC-4.0 (non-commercial) |
precision |
Depth Anything V2 Large | CC-BY-NC-4.0 (non-commercial) |
Use fast for commercial work — it is also the quickest and works very well in practice.
python test_pipeline.pyRuns a synthetic clip through every output format with a stub depth model (no download needed).
- Diffusion-based occlusion inpainting (StereoCrafter-style) to replace the naive warp at disocclusions
- Native MV-HEVC export via AVFoundation (currently delegated to the
spatialCLI)
MIT