mad-bot/ommatidia

Ommatidium dataset

Synthetic renderer-native training and validation data for Ommatidium, generated with Blade.

• Checkpoint: https://huggingface.co/mad-bot/ommatidia
• Format and training code: https://github.com/kvark/ommatidia
• Generator: the ommatidia-data workspace crate

Contents

File	Input source	Samples	Purpose
data/blade-restir/train.omd	raw Blade ReSTIR	2,400	2,040 train / 360 held out
benchmarks/matched-raw.omd	raw Blade ReSTIR	128	matched replacement evaluation
benchmarks/matched-svgf.omd	Blade variance-guided SVGF	128	matched Blade baseline
data/blade-path-trace/1spp-train.omd	sparse path trace, 1 spp	2,400	path-first spatial training
benchmarks/path-trace-1spp-validation.omd	sparse path trace, 1 spp	128	independent path validation
ablations/path-trace-4spp-static-train.omd	sparse path trace, 4 static spp	2,400	favorable history-evidence proxy
benchmarks/path-trace-4spp-static-validation.omd	sparse path trace, 4 static spp	128	independent proxy validation
benchmarks/path-reference-svgf-validation.omd	Blade ReSTIR+SVGF	128	same new canonical references

Every sample contains low-resolution radiance, depth, world-space normals, diffuse albedo, specular F0, and roughness, plus a high-resolution canonical path-traced reference. The primary set uses 128×128 inputs and 256×256 references. Dataset-format version 2 records whether the low-resolution source is raw ReSTIR or SVGF-filtered.

Each benchmark family uses the same 128 procedural scenes and seed 10000; its canonical record payloads are byte-identical. The path-first references use 4,096 spp, eight-bounce targets. The 1/4-spp and SVGF files differ only in the low-resolution evidence paired with those records. The 4-spp files accumulate four samples at a static camera; they are an ablation, not motion-aware temporal training data.

Loading

.omd is Ommatidium's little-endian, contiguous f16 training format. It is designed for the Rust trainer's direct batch reads rather than the Hugging Face Dataset Viewer, which is disabled for this custom binary representation.

hf download mad-bot/ommatidia \
  data/blade-restir/train.omd \
  --repo-type dataset --revision v0.1.0 --local-dir data/hf

cargo run --release -p ommatidia-train -- \
  --data data/hf/data/blade-restir/train.omd \
  --steps 20000 --batch 8 --out runs/ommatidia

See docs/design.md for the semantic contract and the source tree for the authoritative parser.

Generation provenance

• Ommatidium: 7f08f025a3150c355643af513bc2825d88441520
• Blade upstream integration: 3a8895a
• Meganeura upstream integration: 256b906
• Primary generation seed: 1
• Matched validation seed: 10000

Future render estimators belong in this repository as separately identified sources rather than changing the dataset's identity. Training recipes should pin a Hub revision and enumerate the source names they consume.

The first spatial path experiment is intentionally retained even though it is a negative result. On the independent set, the 1-spp model gains only 0.22 dB and the 4-spp model only 0.08 dB. See the full result; these files are the reproducible baseline for the temporal model, not a claim that the current spatial checkpoint solves sparse-path reconstruction.

Limitations

The repository contains small procedural scenes from one renderer and a narrow resolution distribution. It should not be treated as representative of production game content. The canonical targets reduce Monte Carlo noise but are finite-sample path traces rather than analytic ground truth. No uploaded path dataset contains motion vectors, animation, or disocclusions yet.