Ommatidium
Ommatidium is a portable neural denoiser and 2× upscaler for real-time ray tracing. This first checkpoint replaces Blade's variance-guided SVGF pass: it accepts raw ReSTIR radiance and the renderer's depth, normal, diffuse albedo, specular F0, and roughness buffers, then reconstructs a high-resolution frame through Meganeura.
- Source and integration: https://github.com/kvark/ommatidia
- Training and validation data: https://huggingface.co/datasets/mad-bot/ommatidia
- Runtime: Meganeura on a graphics context shared with Blade
- Backends: Vulkan and Metal; the published performance result is Vulkan
Checkpoint
The release contains model.safetensors, config.ron, and a machine-readable
manifest.json. The configuration is a direct, single-frame base-24 U-Net with
three resolution levels, one residual block per level, 649,200 parameters, and
a 2× output scale. It has no temporal history yet.
The input contract is six low-resolution plane groups in this order:
- RGB radiance
- depth
- world-space normal
- diffuse albedo
- specular F0
- roughness
Use the Ommatidium loader rather than constructing the tensor manually: it applies the checkpoint's stored radiance compression and residual gain and keeps the prediction on the shared GPU.
Results
On the matched 128-scene validation capture:
| Reconstruction | Error |
|---|---|
| Raw ReSTIR, nearest 2× | 0.005748 |
| Blade SVGF, nearest 2× | 0.004284 |
| Ommatidium from raw ReSTIR | 0.002876 |
That is 3.01 dB over raw nearest reconstruction and 1.73 dB over the Blade filter it replaces. The canonical references are byte-identical between the raw and SVGF captures.
At the actual rectangular deployment shape, repeated sustained benchmarks measure 19.54–20.91 ms for 960×540 input / 1920×1080 output on an idle Radeon RX 7900 XT with RADV. This includes Ommatidium's texture pack, the 104.1 GFLOP Meganeura network, unpack, and queue submissions; the ray tracer and display post-processing are not included. The benchmark also reports tail latency; a recent 40-frame run measured a 29.95 ms p90.
Intended use and limitations
This is an early research checkpoint intended for Blade integration and for developing portable neural reconstruction runtimes. It was trained entirely on procedural Blade scenes at 128×128 input and 256×256 reference resolution. It may fail on geometry, materials, lighting distributions, resolutions, or renderers outside that training distribution. It is spatial-only, so it cannot recover information from prior frames or guarantee temporal stability.
Pin the v0.1.0 Hub revision, or its exact commit, in applications. Do not
download mutable main for a shipped build.
Revisions
- Ommatidium:
7f08f025a3150c355643af513bc2825d88441520 - Meganeura upstream integration:
256b906 - Blade upstream integration:
3a8895a - Dataset:
mad-bot/ommatidia@v0.1.0
The weights are released under the MIT license.