sharp-android-onnx-test

Experimental SHARP 1024 Android ONNX (folded)

For testing only.

This is an experimental local adaptation of the released Apple SHARP checkpoint sharp_2572gikvuh.pt. It is not a model trained, validated, or released by Apple.

• Contract: sharp-fixed-1024-v1
• Input: float32 image [1, 3, 1024, 1024] and disparity factor [1]
• Output: 524,288 Gaussians (2 x 512 x 512)
• Precision: mixed FP16 model body with FP32 boundaries

Download bundle

Three-stage graph-only pipeline patch

• sharp-android-pipeline-1024-3stage-v1.zip
  • bytes: 38,807,791
  • SHA-256: dbc4ed059b7577760a5bfcf12414e766438c0429196be1910548c243e2b58053

For testing only. This ZIP has exactly four uncompressed entries: its manifest first, followed by stage 1, stage 2, and stage 3 ONNX graphs. It does not contain the 1.306 GB sidecar. The manifest requires the existing sharp-adapted-1024-mixed-fp16-folded.onnx.data file (1,306,759,168 bytes, SHA-256 2b2dcb725647381ff82efc9a2e90dfddf4d0e6ad73d32729617dfcf7dc693c18).

The graphs replace 10 shared residual Relu nodes with Max(x, fp16 zero) and passed the full ONNX checker. On host ORT 1.25 CPU with optimizations disabled, all 12 stage-boundary tensors matched the monolithic graph byte-for-byte, but the five final tensors did not; this patch is not byte-identical to the monolithic execution. A packed-scene point diagnostic on one photo retained the same 130,818 source records and measured about 31.2 dB PSNR, with rare larger attribute outliers, so that result is conditional rather than exact-quality proof.

An SM-S721N probe using stock ORT 1.25, XNNPACK, and NO_OPT ran all three stages in 98.118 seconds with correct output shapes and only finite values. The probe used a deterministic gradient input; it proves runtime feasibility, not monolithic equivalence or complete app visual quality.

• sharp-android-model-1024-folded-conv-output-names-v2.zip
  • bytes: 1,346,019,663
  • SHA-256: 4a41dc6359e9a78b7e829e80f852fa2dfcc0b2e59a4f0fa4d08d56dd50d66242
  • current metadata-only XNNPACK diagnostic-name experiment
• sharp-android-model-1024-folded.zip
  • bytes: 1,346,019,780
  • SHA-256: 1d8433d40ad9183704254bd0c648f30e7db7b9d2d2b39de9098f7ac93eb93552
  • previous folded 1024 graph

Each bundle contains exactly one manifest and its two ONNX artifacts, stored without ZIP compression.

Graph-only patch for an existing v1 sidecar

• sharp-adapted-1024-mixed-fp16-folded-conv-output-names.onnx
  • bytes: 39,258,704
  • SHA-256: cbb29c8e4df76e18e42618fc21eeac3a8e87450c7fef2d294531ebd193102dbb

This standalone graph is a delta-migration artifact for an installation that already has the previous v1 sidecar named sharp-adapted-1024-mixed-fp16-folded.onnx.data (SHA-256 2b2dcb725647381ff82efc9a2e90dfddf4d0e6ad73d32729617dfcf7dc693c18). Place the graph and that existing sidecar in the same directory; the graph's external-data metadata intentionally refers to the existing v1 sidecar name. Do not download or duplicate the 1.306 GB sidecar when that exact hash is already present.

This graph-only patch has the same metadata-only 84-Conv name alignment as the v2 bundle. Model math and weight bytes are unchanged, and Android ARM64 remains unverified.

Adaptation and graph optimization

The local 1024 adaptation changes the two ViT positional grids from 24x24 to 16x16 using bicubic antialiased resampling, changes the SPN patch size from 384 to 256, and remaps SPN merge padding from 3/6 to 2/4. It was not retrained.

The folded graph additionally replaces 192 constant-weight Transpose nodes with pre-transposed MatMul initializers. This offline transformation is mathematically lossless: it changes constant storage layout and graph structure, not the intended model computation.

Conv diagnostic-name alignment candidate

The v2 candidate changes only the diagnostic NodeProto.name fields of 84 Conv nodes. Each name is aligned with that node's existing first output tensor name (for example, node_conv2d_66 becomes conv2d_66). This makes every such diagnostic name also identify a valid produced tensor for an XNNPACK graph-rewrite experiment.

No input or output edge, operator attribute, tensor shape, initializer, model calculation, or weight byte is changed. The external-data artifact is byte-for-byte identical to the previous folded 1024 release.

The exact v2 graph/data pair passed the full ONNX checker, an ONNX Runtime 1.25 CPU ORT_ENABLE_ALL session, and an Android x86_64 ONNX Runtime 1.25 XNNPACK ALL_OPT session-creation smoke test. Android ARM64 has not been verified, so this remains an experimental candidate rather than evidence that the ARM-specific failure is fixed.

Reviewed artifacts

Current v2 candidate (internal bundle names):

• sharp-adapted-1024-mixed-fp16-folded.onnx
  • bytes: 39,258,704
  • SHA-256: cbb29c8e4df76e18e42618fc21eeac3a8e87450c7fef2d294531ebd193102dbb
• sharp-adapted-1024-mixed-fp16-folded.onnx.data
  • bytes: 1,306,759,168
  • SHA-256: 2b2dcb725647381ff82efc9a2e90dfddf4d0e6ad73d32729617dfcf7dc693c18

Previous folded graph:

• sharp-adapted-1024-mixed-fp16-folded.onnx
  • bytes: 39,259,124
  • SHA-256: 05796cd43b94e214fba8d355de6fa7e269d373fce98c7c4d4548b9771315bb08
• sharp-adapted-1024-mixed-fp16-folded.onnx.data
  • bytes: 1,306,759,168
  • SHA-256: 2b2dcb725647381ff82efc9a2e90dfddf4d0e6ad73d32729617dfcf7dc693c18

The adaptation and graph folding do not replace or relax the released checkpoint's license terms. Review those terms before redistribution or use.