xcc.es
[READ-ONLY] Mirror of https://github.com/mrgnw/spatial-maker.
v0.1.1: Default to Small model, update README with model comparison
author
Morgan Williams
date
(Feb 14, 2026, 1:45 PM +0100)
commit
a39d5e1e
a39d5e1ef6fa508024acb7cbf6feb781cfeb5728
parent
9b7682d6
9b7682d67b896b5c2af19eb24b3905e6ae45c612
+387
-55
+11
CLAUDE.md
+11
CLAUDE.md
···
1
+
# Code
2
+
3
+
- Code should be easy to read and understand. Not with comments, but with the coding logic itself.
4
+
- Create modular code for easy readability and maintainability, but do not overabstract. If we're just doing something once or twice, just do things the direct way.
5
+
- Examples should be easy to copy and paste.
6
+
- Dummy variables: Do not _ever_ use fake placeholders like /path/to/your/file or foo.txt. Create a variable with a name and reference that variable in the script or code, so I can then
7
+
- style: use tabs, not spaces
8
+
- Do not add comments or write extra documentation unless I tell you to.
9
+
10
+
# Explanations
11
+
Only explain things or create summaries when I ask you to. Don't write extra documentation unless I tell you to.
+1
-1
Cargo.lock
+1
-1
Cargo.lock
+1
-1
Cargo.toml
+1
-1
Cargo.toml
+118
-52
RUST_README.md
+118
-52
RUST_README.md
···
1
-
# spatial-maker (Rust)
1
+
# spatial-maker
2
2
3
3
Convert 2D images and videos to stereoscopic 3D spatial content for Apple Vision Pro using AI depth estimation.
4
4
5
-
This is a Rust library that uses **Depth Anything V2** via ONNX Runtime to estimate depth from 2D images and create side-by-side (SBS) stereo pairs for 3D viewing.
5
+
[](https://crates.io/crates/spatial-maker)
6
+
[](https://docs.rs/spatial-maker)
6
7
7
8
## Features
8
9
9
-
- ✅ **Fast depth estimation** using ONNX Runtime with CoreML acceleration on Apple Silicon
10
-
- ✅ **High quality stereo generation** using Depth-Image Based Rendering (DIBR)
11
-
- ✅ **Photo support** — convert single images to SBS stereo
12
-
- 🚧 **Video support** — coming soon (per-frame processing with optional temporal smoothing)
13
-
- 🚧 **Video Depth Anything integration** — temporally consistent depth for flicker-free video
10
+
- **Fast depth estimation** — CoreML on Apple Silicon (128ms/frame on M4 Pro) with ONNX fallback
11
+
- **High-quality stereo generation** — Depth-Image Based Rendering (DIBR) with hole filling
12
+
- **Photo & video support** — Process single images or full videos with progress callbacks
13
+
- **Multi-format input** — JPEG, PNG, AVIF, JPEG XL, HEIC via native decoders or ffmpeg
14
+
- **MV-HEVC output** — Side-by-side, top-and-bottom, or separate stereo pairs with optional MV-HEVC packaging
14
15
15
16
## Quick Start
16
17
17
-
### Download the Model
18
+
Add to your `Cargo.toml`:
19
+
20
+
```toml
21
+
[dependencies]
22
+
spatial-maker = "0.1"
23
+
```
24
+
25
+
### Process a Photo
18
26
19
-
Download the Depth Anything V2 Small ONNX model:
27
+
```rust
28
+
use spatial_maker::{process_photo, SpatialConfig, OutputOptions};
20
29
21
-
```bash
22
-
mkdir -p ~/.spatial-maker/checkpoints
23
-
curl -L -o ~/.spatial-maker/checkpoints/depth_anything_v2_vits.onnx \
24
-
"https://github.com/fabio-sim/Depth-Anything-ONNX/releases/download/v2.0.0/depth_anything_v2_vits.onnx"
25
-
```
30
+
#[tokio::main]
31
+
async fn main() -> Result<(), Box<dyn std::error::Error>> {
32
+
let config = SpatialConfig::default(); // Uses Small model (48MB, auto-downloaded)
33
+
let output_opts = OutputOptions::default(); // Side-by-side JPEG
26
34
27
-
### Usage (Photo Example)
35
+
process_photo(
36
+
"input.jpg".as_ref(),
37
+
"output_sbs.jpg".as_ref(),
38
+
config,
39
+
output_opts,
40
+
).await?;
28
41
29
-
```bash
30
-
cargo run --example photo --release -- input.jpg -o output_sbs.jpg --max-disparity 30
42
+
Ok(())
43
+
}
31
44
```
32
45
33
-
Or use as a library:
46
+
### Process a Video
34
47
35
48
```rust
36
-
use spatial_maker::process_photo;
37
-
use image::open;
49
+
use spatial_maker::{process_video, SpatialConfig, VideoProgress};
38
50
39
-
fn main() -> anyhow::Result<()> {
40
-
let input = open("photo.jpg")?;
41
-
let sbs = process_photo(
42
-
input,
43
-
"~/.spatial-maker/checkpoints/depth_anything_v2_vits.onnx",
44
-
30.0 // max disparity in pixels
45
-
)?;
46
-
sbs.save("photo_sbs.jpg")?;
51
+
#[tokio::main]
52
+
async fn main() -> Result<(), Box<dyn std::error::Error>> {
53
+
let config = SpatialConfig::default();
54
+
55
+
process_video(
56
+
"input.mp4".as_ref(),
57
+
"output_sbs.mp4".as_ref(),
58
+
config,
59
+
Some(Box::new(|progress: VideoProgress| {
60
+
println!("{}: {:.1}%", progress.stage, progress.percent);
61
+
})),
62
+
).await?;
63
+
47
64
Ok(())
48
65
}
49
66
```
50
67
68
+
## Model Sizes
69
+
70
+
Models are auto-downloaded from [HuggingFace](https://huggingface.co/mrgnw/depth-anything-v2-coreml) on first use:
71
+
72
+
| Size | Download | Quality | Speed (M4 Pro) | License |
73
+
|------|----------|---------|----------------|---------|
74
+
| **Small** (default) | 48 MB | Good | ~85ms/frame | Apache-2.0 ✅ |
75
+
| **Base** | 186 MB | Better | ~128ms/frame | CC-BY-NC-4.0 ⚠️ |
76
+
| **Large** | 638 MB | Best | ~190ms/frame | CC-BY-NC-4.0 ⚠️ |
77
+
78
+
⚠️ Base and Large models are **non-commercial only** (CC-BY-NC-4.0). Small is Apache-2.0 (commercial OK).
79
+
80
+
Select a model:
81
+
82
+
```rust
83
+
let config = SpatialConfig {
84
+
encoder_size: "b".to_string(), // "s", "b", or "l"
85
+
max_disparity: 30,
86
+
..Default::default()
87
+
};
88
+
```
89
+
90
+
## Feature Flags
91
+
92
+
```toml
93
+
[dependencies]
94
+
spatial-maker = { version = "0.1", features = ["onnx"] }
95
+
```
96
+
97
+
| Feature | Default | Description |
98
+
|---------|---------|-------------|
99
+
| `coreml` | ✅ | CoreML via Swift FFI (macOS only, ~3x faster than ONNX CPU) |
100
+
| `onnx` | ❌ | ONNX Runtime fallback (cross-platform) |
101
+
| `avif` | ❌ | Native AVIF decoder (requires system libdav1d) |
102
+
| `jxl` | ❌ | Native JPEG XL decoder (pure Rust via jxl-oxide) |
103
+
| `heic` | ❌ | Native HEIC decoder (requires system libheif) |
104
+
105
+
Formats not enabled via feature flags fall back to ffmpeg conversion (slower but works for everything).
106
+
51
107
## How It Works
52
108
53
-
1. **Depth Estimation**: Uses Depth Anything V2 (24.8M params) via ONNX Runtime to estimate per-pixel depth
54
-
2. **Preprocessing**: Resizes image to 518x518, normalizes with ImageNet mean/std
55
-
3. **Stereo Generation**: Uses DIBR to create left/right eye views by shifting pixels based on depth
56
-
4. **Output**: Horizontally stacks left and right views into a side-by-side image
109
+
1. **Depth Estimation**: [Depth Anything V2](https://github.com/DepthAnything/Depth-Anything-V2) via CoreML (default) or ONNX
110
+
2. **Preprocessing**: Resize to 518×518, normalize with ImageNet stats, NCHW tensor format
111
+
3. **Stereo Generation**: DIBR shifts pixels based on depth; expanding-ring hole filling for disocclusions
112
+
4. **Video Pipeline**: `ffmpeg` frame extraction → depth+stereo (parallelized) → `ffmpeg` encoding with model caching
113
+
114
+
## API Overview
57
115
58
-
## Model Info
116
+
```rust
117
+
// High-level API (most common)
118
+
pub async fn process_photo(input, output, config, output_options) -> SpatialResult<()>
119
+
pub async fn process_video(input, output, config, progress_cb) -> SpatialResult<()>
120
+
121
+
// Low-level API (custom pipelines)
122
+
pub struct CoreMLDepthEstimator; // macOS CoreML backend
123
+
pub struct OnnxDepthEstimator; // ONNX Runtime backend
124
+
pub fn generate_stereo_pair(image, depth, max_disparity) -> SpatialResult<(DynamicImage, DynamicImage)>
125
+
pub fn create_sbs_image(left, right) -> DynamicImage
126
+
```
59
127
60
-
- **Model**: [Depth Anything V2 Small](https://github.com/DepthAnything/Depth-Anything-V2)
61
-
- **ONNX Export**: [fabio-sim/Depth-Anything-ONNX](https://github.com/fabio-sim/Depth-Anything-ONNX)
62
-
- **License**: Apache-2.0
63
-
- **Size**: 95 MB
64
-
- **Speed**: ~13ms on RTX 4080, faster on Apple Silicon with CoreML
128
+
See [docs.rs/spatial-maker](https://docs.rs/spatial-maker) for full API documentation.
65
129
66
-
## Roadmap
130
+
## Requirements
67
131
68
-
- [x] Photo support (SBS output)
69
-
- [x] ONNX model integration
70
-
- [x] Model checkpoint discovery
71
-
- [ ] CoreML execution provider testing
72
-
- [ ] Video frame processing
73
-
- [ ] Video Depth Anything integration (temporal consistency)
74
-
- [ ] Integration with Frame app (replace Python subprocess)
132
+
### For CoreML (default, macOS only):
133
+
- macOS 13.0+ (Ventura or later)
134
+
- Apple Silicon recommended (falls back to CPU on Intel)
135
+
- Xcode Command Line Tools (for Swift compilation during build)
75
136
76
-
## Development
137
+
### For ONNX (optional, cross-platform):
138
+
- Any OS (Linux, Windows, macOS)
139
+
- Binaries auto-downloaded via `ort` crate
77
140
78
-
Built with:
79
-
- [ort](https://github.com/pykeio/ort) — ONNX Runtime bindings for Rust
80
-
- [image](https://github.com/image-rs/image) — Image encoding/decoding
81
-
- [ndarray](https://github.com/rust-ndarray/ndarray) — N-dimensional arrays
141
+
### For Video:
142
+
- `ffmpeg` in PATH — [https://ffmpeg.org](https://ffmpeg.org)
143
+
- `spatial` CLI for MV-HEVC output — [https://blog.mikeswanson.com/spatial](https://blog.mikeswanson.com/spatial)
82
144
83
-
See [research/rust-depth.md](research/rust-depth.md) for detailed design decisions and model comparisons.
145
+
```bash
146
+
brew install ffmpeg spatial
147
+
```
84
148
85
149
## License
86
150
87
151
MIT
152
+
153
+
Models: Small (Apache-2.0), Base/Large (CC-BY-NC-4.0). See [HuggingFace repo](https://huggingface.co/mrgnw/depth-anything-v2-coreml) for details.
+154
scripts/upload_to_huggingface.md
+154
scripts/upload_to_huggingface.md
···
1
+
# Upload CoreML Models to HuggingFace
2
+
3
+
Follow these steps to upload the converted CoreML models to HuggingFace for easy distribution.
4
+
5
+
## 1. Create HuggingFace Repository
6
+
7
+
1. Go to https://huggingface.co/new
8
+
2. Create a new **model** repository
9
+
3. Name it: `depth-anything-v2-coreml` (or your preferred name)
10
+
4. Make it **public** so users can download without authentication
11
+
5. Add a description: "Depth Anything V2 models in CoreML format for Apple Silicon"
12
+
13
+
## 2. Prepare Models for Upload
14
+
15
+
The models are already in `checkpoints/`. We need to compress them:
16
+
17
+
```bash
18
+
cd checkpoints
19
+
20
+
# Compress Base model
21
+
tar -czf DepthAnythingV2BaseF16.mlpackage.tar.gz DepthAnythingV2BaseF16.mlpackage
22
+
23
+
# Compress Large model
24
+
tar -czf DepthAnythingV2LargeF16.mlpackage.tar.gz DepthAnythingV2LargeF16.mlpackage
25
+
26
+
# Check sizes
27
+
ls -lh *.tar.gz
28
+
```
29
+
30
+
This compresses the models from ~870 MB to ~430 MB total.
31
+
32
+
## 3. Upload via Web Interface (Easiest)
33
+
34
+
1. Go to your repo: `https://huggingface.co/YOUR_USERNAME/depth-anything-v2-coreml`
35
+
2. Click "Files and versions" → "Add file" → "Upload files"
36
+
3. Upload:
37
+
- `DepthAnythingV2BaseF16.mlpackage.tar.gz`
38
+
- `DepthAnythingV2LargeF16.mlpackage.tar.gz`
39
+
4. Commit with message: "Add CoreML models (Base and Large)"
40
+
41
+
**Note**: We don't upload Small because it's available directly from Apple.
42
+
43
+
## 4. Upload via Git LFS (Alternative)
44
+
45
+
If you prefer command-line:
46
+
47
+
```bash
48
+
# Install git-lfs if needed
49
+
brew install git-lfs
50
+
git lfs install
51
+
52
+
# Clone your HF repo
53
+
git clone https://huggingface.co/YOUR_USERNAME/depth-anything-v2-coreml
54
+
cd depth-anything-v2-coreml
55
+
56
+
# Track tar.gz files with LFS
57
+
git lfs track "*.tar.gz"
58
+
git add .gitattributes
59
+
60
+
# Copy and commit models
61
+
cp ../spatial-maker/checkpoints/*.tar.gz .
62
+
git add *.tar.gz
63
+
git commit -m "Add CoreML models (Base and Large)"
64
+
git push
65
+
```
66
+
67
+
## 5. Create Model Card (README.md)
68
+
69
+
Add this to your HuggingFace repo's README.md:
70
+
71
+
````markdown
72
+
---
73
+
license: apache-2.0
74
+
tags:
75
+
- depth-estimation
76
+
- coreml
77
+
- apple-silicon
78
+
- vision
79
+
---
80
+
81
+
# Depth Anything V2 - CoreML
82
+
83
+
Depth Anything V2 models converted to CoreML format for optimized inference on Apple Silicon (M-series chips).
84
+
85
+
## Models
86
+
87
+
| Model | Size | Parameters | Performance (M4 Pro) | License |
88
+
|-------|------|------------|----------------------|---------|
89
+
| Small F16 | 48 MB | 24.8M | ~30ms (~33 fps) | Apache-2.0 |
90
+
| Base F16 | 186 MB | 97.5M | ~60-90ms (~14 fps) | Apache-2.0 |
91
+
| Large F16 | 638 MB | 335.3M | ~200-300ms (~4 fps) | CC-BY-NC-4.0 |
92
+
93
+
All models use Float16 precision and run on Apple's Neural Engine + GPU + CPU.
94
+
95
+
## Download
96
+
97
+
```bash
98
+
# Download from spatial-maker repo
99
+
git clone https://github.com/YOUR_USERNAME/spatial-maker
100
+
cd spatial-maker
101
+
./scripts/download_coreml_models.sh
102
+
```
103
+
104
+
Or download directly:
105
+
- [Base model](https://huggingface.co/YOUR_USERNAME/depth-anything-v2-coreml/resolve/main/DepthAnythingV2BaseF16.mlpackage.tar.gz) (186 MB)
106
+
- [Large model](https://huggingface.co/YOUR_USERNAME/depth-anything-v2-coreml/resolve/main/DepthAnythingV2LargeF16.mlpackage.tar.gz) (638 MB)
107
+
108
+
Small model available from [Apple's CoreML model zoo](https://developer.apple.com/machine-learning/models/).
109
+
110
+
## Usage
111
+
112
+
```swift
113
+
import CoreML
114
+
115
+
let model = try MLModel(contentsOf: URL(fileURLWithPath: "DepthAnythingV2BaseF16.mlpackage"))
116
+
// ... inference code
117
+
```
118
+
119
+
## Citation
120
+
121
+
```bibtex
122
+
@article{yang2024depth,
123
+
title={Depth Anything V2},
124
+
author={Yang, Lihe and Kang, Bingyi and Huang, Zilong and Zhao, Zhen and Xu, Xiaogang and Feng, Jiashi and Zhao, Hengshuang},
125
+
journal={arXiv:2406.09414},
126
+
year={2024}
127
+
}
128
+
```
129
+
130
+
## Conversion
131
+
132
+
These models were converted from PyTorch using `coremltools`. See [conversion script](https://github.com/YOUR_USERNAME/spatial-maker/tree/main/scripts/coreml_conversion).
133
+
````
134
+
135
+
## 6. Update Download Script
136
+
137
+
After uploading, update `scripts/download_coreml_models.sh` with your actual HuggingFace username:
138
+
139
+
```bash
140
+
HF_REPO="YOUR_ACTUAL_USERNAME/depth-anything-v2-coreml"
141
+
```
142
+
143
+
## Verification
144
+
145
+
Test the download works:
146
+
147
+
```bash
148
+
cd /tmp
149
+
git clone https://github.com/YOUR_USERNAME/spatial-maker
150
+
cd spatial-maker
151
+
./scripts/download_coreml_models.sh
152
+
```
153
+
154
+
Should download all three models successfully!
+1
-1
src/lib.rs
+1
-1
src/lib.rs
test_baseline.jpg
test_baseline.jpg
This is a binary file and will not be displayed.
test_coreml.jpg
test_coreml.jpg
This is a binary file and will not be displayed.
test_coreml_1.jpg
test_coreml_1.jpg
This is a binary file and will not be displayed.
test_coreml_2.jpg
test_coreml_2.jpg
This is a binary file and will not be displayed.
test_coreml_3.jpg
test_coreml_3.jpg
This is a binary file and will not be displayed.
test_coreml_4.jpg
test_coreml_4.jpg
This is a binary file and will not be displayed.
test_coreml_5.jpg
test_coreml_5.jpg
This is a binary file and will not be displayed.
test_output.jpg
test_output.jpg
This is a binary file and will not be displayed.
test_run1.jpg
test_run1.jpg
This is a binary file and will not be displayed.
test_run2.jpg
test_run2.jpg
This is a binary file and will not be displayed.
test_run3.jpg
test_run3.jpg
This is a binary file and will not be displayed.
test_run4.jpg
test_run4.jpg
This is a binary file and will not be displayed.
test_run5.jpg
test_run5.jpg
This is a binary file and will not be displayed.
+1
todos/_top.md
+1
todos/_top.md
···
1
+
1. spatial-settings-overhaul.md
+100
todos/spatial-settings-overhaul.md
+100
todos/spatial-settings-overhaul.md
···
1
+
# Spatial Settings Panel Overhaul
2
+
3
+
Improve spatial settings panel: model downloading, unified start button, eliminate redundant downscaling.
4
+
5
+
## Status: COMPLETE
6
+
7
+
All tasks done. `svelte-check` passes with 0 errors. `cargo check` passes.
8
+
9
+
## Decisions Made
10
+
11
+
- **Downscale integration:** Two-pass approach. Frame GUI does ffmpeg conversion first (user's resolution/codec/trim settings), then spatial-maker processes the output with `--skip-downscale`.
12
+
- **FPS:** Keep 24fps hardcoded for spatial (Vision Pro preference).
13
+
- **Model download:** Rust/Tauri command with reqwest streaming + progress events to the UI.
14
+
15
+
## Tasks
16
+
17
+
### 1. Rust: Add reqwest dependency and model check/download Tauri commands
18
+
- [x] Add `reqwest` (with `stream` feature) and `futures-util` to Cargo.toml
19
+
- [x] Add `ModelDownloadProgressPayload`, `ModelDownloadCompletePayload`, `ModelDownloadErrorPayload` to `spatial/types.rs`
20
+
- [x] Add `enabled: bool` field to `SpatialConfig` Rust struct
21
+
- [x] Add `check_spatial_models` command (checks `~/.spatial-maker/checkpoints/` for each model)
22
+
- [x] Add `download_spatial_model` command (streams from HuggingFace, emits progress events)
23
+
- Files changed:
24
+
- `frame/src-tauri/Cargo.toml`
25
+
- `frame/src-tauri/src/spatial/types.rs`
26
+
- `frame/src-tauri/src/spatial/commands.rs`
27
+
28
+
### 2. Rust: Register new commands in lib.rs
29
+
- [x] Add `check_spatial_models` and `download_spatial_model` to invoke_handler
30
+
- File: `frame/src-tauri/src/lib.rs`
31
+
32
+
### 3. Svelte: Add 'enabled' field to SpatialConfig type + service layer
33
+
- [x] Add `enabled: boolean` to `SpatialConfig` interface in `types.ts` (default: `false`)
34
+
- [x] Default `skipDownscale` to `true` (no longer user-facing, always skipped)
35
+
- [x] Add `checkSpatialModels()` and `downloadSpatialModel()` to spatial service
36
+
- [x] Add `setupModelDownloadListeners()` for progress/complete/error events
37
+
- Files:
38
+
- `frame/src/lib/types.ts`
39
+
- `frame/src/lib/services/spatial.ts`
40
+
41
+
### 4. Svelte: Overhaul SpatialTab UI
42
+
- [x] Add "Enable spatial encoding" checkbox at the top (prominent)
43
+
- [x] Gray out rest of settings when disabled (opacity + pointer-events-none)
44
+
- [x] Model selector: show download status per model (check icon / arrow icon / progress %)
45
+
- [x] Clicking unavailable model triggers download with inline progress bar
46
+
- [x] Fix spacing issues (tighter gaps, consistent sizing)
47
+
- [x] Remove "Skip downscale" checkbox (always skipped now)
48
+
- [x] Remove old info box about manual installation
49
+
- File: `frame/src/lib/components/settings/tabs/SpatialTab.svelte`
50
+
51
+
### 5. Svelte: Remove Spatial button from titlebar
52
+
- [x] Remove `onStartSpatial` prop and the Spatial button from all 3 titlebars
53
+
- [x] Remove `onStartSpatial` from Titlebar.svelte wrapper
54
+
- [x] Remove `IconGlasses` import from titlebars
55
+
- [x] Clean up +page.svelte (remove onStartSpatial binding)
56
+
- Files:
57
+
- `frame/src/lib/components/layout/titlebar/MacosTitlebar.svelte`
58
+
- `frame/src/lib/components/layout/titlebar/WindowsTitlebar.svelte`
59
+
- `frame/src/lib/components/layout/titlebar/LinuxTitlebar.svelte`
60
+
- `frame/src/lib/components/layout/Titlebar.svelte`
61
+
- `frame/src/routes/+page.svelte`
62
+
63
+
### 6. Svelte: Wire Start button to chain spatial after conversion
64
+
- [x] Add `onConversionCompleted` callback to `ConversionCallbacks` interface
65
+
- [x] In `useConversionQueue.svelte.ts` `onCompleted` handler: call callback instead of marking COMPLETED directly
66
+
- [x] Add `queueSpatialForFile(id, filePath)` method to spatial queue
67
+
- [x] In `+page.svelte`: wire callback to check `spatialQueue.config.enabled` and chain spatial
68
+
- Files:
69
+
- `frame/src/lib/features/conversion/useConversionQueue.svelte.ts`
70
+
- `frame/src/lib/features/spatial/useSpatialQueue.svelte.ts`
71
+
- `frame/src/routes/+page.svelte`
72
+
73
+
### 7. Rust: Always pass --skip-downscale when invoked from Frame
74
+
- [x] In `spatial/worker.rs`, always add `--skip-downscale` flag
75
+
- [x] Removed conditional `if task.config.skip_downscale` check
76
+
- [x] Updated progress mapping (depth_stereo 0-85% instead of 10-85%)
77
+
- File: `frame/src-tauri/src/spatial/worker.rs`
78
+
79
+
### 8. Build and verify compilation
80
+
- [x] `cargo check` passes (1 pre-existing dead_code warning for TaskNotFound)
81
+
- [x] `svelte-check` passes with 0 errors, 0 warnings
82
+
83
+
## Architecture Notes
84
+
85
+
### Flow (after changes)
86
+
```
87
+
User clicks Start
88
+
-> Frame ffmpeg conversion (resolution, codec, trim, etc.)
89
+
-> conversion-completed event (with outputPath)
90
+
-> if spatialConfig.enabled:
91
+
-> queue_spatial(id, outputPath, spatialConfig) with --skip-downscale
92
+
-> spatial-maker: depth+stereo -> audio mux -> MV-HEVC
93
+
-> spatial-completed event
94
+
-> Mark file as COMPLETED
95
+
```
96
+
97
+
### Model URLs
98
+
- Small (vits, ~99MB): https://huggingface.co/depth-anything/Depth-Anything-V2-Small/resolve/main/depth_anything_v2_vits.pth
99
+
- Base (vitb, ~390MB): https://huggingface.co/depth-anything/Depth-Anything-V2-Base/resolve/main/depth_anything_v2_vitb.pth
100
+
- Large (vitl, ~1.3GB): https://huggingface.co/depth-anything/Depth-Anything-V2-Large/resolve/main/depth_anything_v2_vitl.pth