| |
| |
| |
|
|
| \subsection{End-to-end case study: DinoVision} |
| \label{sec:dinovision} |
|
|
| We evaluate Meganeura beyond isolated operators with DinoVision, an Android XR |
| application that reconstructs passthrough imagery from intermediate |
| DINOv3~\cite{simeoni2025dinov3} features. The application center-crops and |
| resizes an RGB frame to |
| $224\times224$, executes the first three layers of a frozen DINOv3 ViT-S/16 |
| encoder, rearranges its $14\times14\times384$ patch features, and decodes them |
| to RGB. The encoder and decoder form one batch-one inference graph on the |
| device. In the live application, Meganeura sessions share Blade's Vulkan |
| context and queue with OpenXR rendering. |
|
|
| This case study exercises patch projection, learned prefix tokens, axial RoPE, |
| multi-head attention, LayerNorm, GELU MLPs, LayerScale, residual connections, |
| convolution, group normalization, SiLU, upsampling, autodiff, Adam, parameter |
| serialization, and Android cross-compilation. It is evidence about the |
| breadth and integration of the Meganeura stack, rather than a comparison with |
| PyTorch or a claim of capture-to-photon latency. |
|
|
| \paragraph{Training and evaluation.} |
| The frozen encoder is used to cache features on an NVIDIA host, after which |
| Meganeura trains a 2,012,547-parameter decoder-only batch graph. Deployment |
| uses the same decoder construction and learned parameters in a batch-one |
| joined graph. Thus the two paths share IR operations, compiler, and runtime, |
| but are not literally the same complete graph. |
|
|
| We preserve Imagenette's upstream train/validation split. Each of three runs |
| uses the same class-interleaved 2,500-image training subset (250 images per |
| class), 12,000 batch-eight Adam updates, and a distinct seed controlling both |
| initialization and data order. We evaluate every run on all 3,925 validation |
| images. Seed zero was selected for deployment before validation; seed one is |
| the strongest replicate, ruling out best-seed selection of the deployed |
| artifact. Table~\ref{tab:dinovision-quality} reports held-out quality. The |
| aggregate covers 11,775 reconstruction executions, with the same validation |
| set reused across seeds. |
|
|
| \begin{table}[t] |
| \centering |
| \caption{Held-out DinoVision reconstruction quality. The last row is the |
| mean $\pm$ sample standard deviation across independently initialized |
| decoders.} |
| \label{tab:dinovision-quality} |
| \begin{tabular}{rcccc} |
| \hline |
| Seed & Global PSNR & Median PSNR & Median SSIM & Median MAE \\ |
| & (dB) & (dB) & & \\ |
| \hline |
| 0 (deployed) & 21.7688 & 22.3603 & 0.6338 & 0.04729 \\ |
| 1 & 21.9830 & 22.5714 & 0.6491 & 0.04598 \\ |
| 2 & 21.8170 & 22.3911 & 0.6388 & 0.04700 \\ |
| mean $\pm$ SD & $21.86\pm0.11$ & $22.44\pm0.11$ & |
| $0.6406\pm0.0078$ & $0.0468\pm0.0007$ \\ |
| \hline |
| \end{tabular} |
| \end{table} |
|
|
| \paragraph{Independent correctness gate.} |
| The audit found a silent layout error in the initial implementation: |
| Meganeura's NCHW-flat per-channel multiplication was applied directly to a |
| $[\mathrm{tokens},\mathrm{hidden}]$ LayerScale branch. Only the first token |
| received the intended residual update, yet downstream images remained |
| plausible. We invalidated all weights and measurements produced by that |
| graph, corrected the layout, and retrained all three replicates. |
|
|
| For the corrected implementation, an independent Torch/Transformers reference |
| and Meganeura consume the same normalized floating-point tensor and checkpoint |
| parameters. The predeclared acceptance thresholds are relative $L_2\leq |
| 0.01$, CLS cosine $>0.999$, and every patch-token cosine $>0.999$. |
| Table~\ref{tab:dinovision-correctness} shows that the deployed three-layer |
| graph and the complete twelve-layer control pass. Exact-shape CPU tests also |
| cover the attention and LayerScale operators. |
|
|
| \begin{table}[t] |
| \centering |
| \caption{Meganeura encoder agreement with the independent reference.} |
| \label{tab:dinovision-correctness} |
| \begin{tabular}{rccc} |
| \hline |
| Encoder depth & Relative $L_2$ & CLS cosine & Worst patch cosine \\ |
| \hline |
| 1 & 0.000781 & 1.000000 & 0.999990 \\ |
| 3 & 0.001403 & 1.000000 & 0.999995 \\ |
| 12 & 0.002253 & 0.999997 & 0.999996 \\ |
| \hline |
| \end{tabular} |
| \end{table} |
|
|
| The failure suggests a reusable Meganeura hardening change: graph construction |
| should validate the rank, flattened element count, channel/spatial product, |
| and gate length of per-channel multiplication. The compiler revision used |
| here derives dispatch length from the first shape coordinate, so the invalid |
| layout otherwise fails silently. |
|
|
| \paragraph{Deployment gate.} |
| The corrected graph cross-compiles for Android API 34/aarch64, including the |
| native evaluator, benchmark, and XR APK. A fixed-frame comparison on Quest 3S |
| passes before timing: patches are bit-identical; encoder output has relative |
| $L_2=0.001059$, cosine 0.99999944, and minimum token cosine 0.999973; spatial |
| decoder input has relative $L_2=0.001063$; and reconstruction has relative |
| $L_2=0.000450$ and cosine 0.99999994. We do not report historical headset |
| numbers: they were generated by the invalid graph and a minimum-of-samples |
| timing protocol. Device performance is admissible only after (1) three fresh |
| native processes each retain at least 20 synchronized samples after warmup; |
| and (2) a live-worn XR sweep records all predeclared submission-chunk settings |
| with power, thermal, display, and runtime state. Until those timing gates pass, |
| DinoVision supports Android numerical-correctness but not device-performance |
| claims. |
|
|
| The public artifact at |
| \texttt{https://huggingface.co/mad-bot/dinovision} contains the selected and |
| replicate decoder weights, per-image held-out records, immutable manifests, |
| independent-reference tensors, and exact source snapshots. It deliberately |
| excludes the gated DINOv3 checkpoint, dataset photographs, and private headset |
| captures. |
|
|
