DinoVision decoder

Built with DINOv3

This repository distributes the decoder weights and audited evidence for DinoVision, a train-to-deploy case study for Meganeura. Meganeura trains the RGB decoder on an NVIDIA Vulkan host; the same decoder definition and parameters then run beside a frozen DINOv3 encoder in a Blade/OpenXR application on a Quest 3S.

This is an artifact repository, not a standalone Hugging Face pipeline. The GitHub repository owns source and developer instructions. Its paper-facing report owns interpretation. This card indexes the weights, exact provenance, and machine-readable evidence without copying the report.

Qualitative headset capture

Prompt

Quest 3S live reconstruction capture (39 seconds)

Prompt

Quest 3S reconstruction screenshot 1

Prompt

Quest 3S reconstruction screenshot 2

Prompt

Quest 3S reconstruction screenshot 3

These Meta Quest 3S casting captures show the live reconstruction output. They are qualitative project media, not additional benchmark observations or capture-to-photon evidence. File metadata and hashes are recorded in media/metadata.json.

The quantitative artifact and its unchanged ARTIFACT_MANIFEST.json are frozen at evidence-v1. The media above was added afterward and is intentionally outside that manifest.

Required base model

The gated encoder is deliberately not redistributed. Obtain facebook/dinov3-vits16-pretrain-lvd1689m after accepting Meta's terms.

Property	Value
Encoder checkpoint SHA-256	4610ad75edef83e75afdebf162d148dc628045ea6cbb83d67d4708c709c4f91d
Encoder used by DinoVision	first 3 of 12 ViT-S/16 layers
Input / output	normalized RGB 224 x 224 / sigmoid RGB 224 x 224
Patch representation	14 x 14 x 384
Decoder parameters	2,012,547

Selected weights

Seed 0 was selected for deployment before validation; it was not chosen as the best of the three runs.

File	SHA-256	Purpose
decoder.bin	dcc2b9560fbe0d59869ea2e37d900c1b72dbf01631b5869fd0bae6ca16dff846	little-endian f32 runtime parameters in graph declaration order
decoder.safetensors	ead276f10ad69f417ebe0002152f4853889ac7c70c37251b43c34a3f7462b262	the same values as 26 named tensors with provenance metadata

The SafeTensors conversion verifies exact byte reconstruction of the runtime format. Replicate weights are under replicates/; the generated held-out summary is quality/summary.md, with all 3,925 per-image records retained for each seed.

Training cell

Each replicate uses the same class-interleaved 2,500-image subset of the Imagenette training split, 12,000 batch-eight updates, mean absolute error, and Adam with a linear learning-rate decay from 0.002 to a 0.0001 floor. Seeds independently control initialization and data order. Evaluation covers the complete upstream 3,925-image validation split and records MSE, MAE, PSNR, and RGB SSIM per image.

Training wall time is excluded from performance claims because the host was interactive and experienced recorded suspensions. run-notes.json preserves that disposition for each replicate.

Evidence map

Path	Evidence
quality/	per-seed held-out records and independently generated aggregate
manifests/	immutable dataset and fixed-frame manifests plus summaries
correctness/reference/	Torch/Transformers tensors and Meganeura stage-by-stage comparisons
correctness/host-quest/	fixed-frame RTX/Adreno tensor comparisons
benchmarks/isolated/	three-process synchronized Quest samples and state metadata
benchmarks/live-xr/	live-worn submission-chunk records and co-tenancy summary
environment/	source revisions, exact training snapshot, Cargo resolution, toolchains, and checked Android build
reproducibility/	short byte-identical seeded replay records

ARTIFACT_MANIFEST.json hashes every payload. The repository's validator recomputes aggregate distributions from raw records, binds device binaries to the checked Android build, checks state and protocol invariants, verifies all source/weight hashes, and rejects unrecognized model or image payloads.

Dataset photographs, private headset captures, the gated DINOv3 checkpoint, and host/device identifiers are not included. The one public correctness frame is synthetic and hash-bound.

Intended use and limitations

These weights support reproduction of a systems case study. They are not intended for photographic restoration, surveillance, identity inference, medical use, or safety-critical perception. Quality is measured on held-out photographs rather than independent headset sessions. Device worker latency is not capture-to-photon latency, and the artifact establishes no PyTorch-on-Android speedup or on-device training result.

License and acknowledgment

DinoVision source is MIT-licensed. The decoder weights are treated conservatively as DINOv3-derived and distributed under the included DINOv3-LICENSE.md; see NOTICE.md. Use of the required base model and these weights is subject to that agreement. Research using this artifact must acknowledge DINOv3 and should consult the live Meta terms because they may be amended.

@misc{simeoni2025dinov3,
  title         = {{DINOv3}},
  author        = {Sim{\'e}oni, Oriane and Vo, Huy V. and Seitzer, Maximilian and others},
  year          = {2025},
  eprint        = {2508.10104},
  archivePrefix = {arXiv},
  primaryClass  = {cs.CV},
  url           = {https://arxiv.org/abs/2508.10104}
}