runanywhere/nemotron-ocr-v1-full_ANE

Nemotron-OCR v1 — Apple Neural Engine (detector + recognizer)

The complete OCR pipeline for the ANE: a FOTS/RegNet text detector together with the CTC line recognizer, as one NeuRT bundle.

Why this exists

The previously published nemotron-ocr-*_ANE bundles carry only the recognizer. That is not a runnable OCR model, and the reason is easy to miss: the recognizer's input is crop [1, 128, 8, 32] — a grid-sampled crop of the detector's 128-channel feature map, not pixels. With no detector, nothing can produce a single input tensor. The upstream recipe says so directly: "a real input is a grid-sampled crop of the detector's feature map, which needs the RegNet forward plus the C++ IndirectGridSample."

So a recognizer-only bundle loads and is undrivable. The RunAnywhere runtime now refuses one by name rather than accepting it.

Graphs

role	shape
detector	image [1,3,960,960] → confidence [1,240,240], rbox [1,240,240,5], feature map [1,128,240,240]
recognizer	crop [1,128,8,32] → CTC logits [1,1,32,858]

The detector's third output is the 128 channels the recognizer consumes — that is what makes the pair a pipeline rather than two models.

Measured on an M4 Max

Gate A, against the torch reference on the package's own example image:

output	cosine
confidence	0.99995617
rbox_geometry	0.99993401
feature map	0.99999610

min cosine 0.99993401 against a 0.999 bar. Detector latency cpuAndNeuralEngine 89.55 ms vs cpuOnly 186.65 ms — 2.08× faster with the ANE admitted, which is what genuine residency looks like.

Host-side stages

Everything ANE-hostile is post-processing and stays on the host, exactly as upstream splits it: NMS, rrect_to_quads, quad_rectify, indirect_grid_sample, and CTC collapse. The graphs themselves are conv/BN/ReLU plus squeeze-excite — the friendliest shape the ANE has.

Output tensors are named var_NNNN by the converter and must be read by declared name, never position; *_detector.iodesc.json records the mapping.