alastorid/TruFor-DnCNN-CoreML-FP16

TruFor DnCNN Core ML FP16

This repository contains an FP16-compute Core ML conversion of only TruFor's 17-convolution DnCNN Noiseprint++ subnetwork. It is not a complete TruFor model. The full CMX encoder, localization/confidence heads, and trained detection head are not in this package. WaifuFinder runs those remaining stages with its native Swift/MLX FP32 implementation.

The graph uses FP16 weights and internal activations where supported by Core ML; its public RGB and validity-mask inputs and noiseprint output are FP32. The fixed input shape is [1, 3, 546, 546] for RGB plus [1, 1, 546, 546] for the validity mask. Inputs use opaque sRGB RGB divided by 256. A 17-pixel halo surrounds each 512-pixel output tile. The portable artifact is model.mlpackage; compile it locally on the target Mac before use. The inference runtime is Apple Core ML and native Swift/Objective-C; no Python is in the inference path.

M4 measurements

Measurements use the same verified weights and 2000×2000 fixture on an Apple M4 running macOS 26.4.1. The FP16 DnCNN tile was compared with the native MLX FP32 reference over the central 512×512 pixels: maximum absolute error 0.2437461019 and relative RMS error 0.5971685071. The FP32 Core ML counterpart measured 0.0004192144 maximum absolute error and 0.0007005027 relative RMS error.

In the full 4 MP TruFor pipeline, with this Core ML FP16 DnCNN output and the remaining stages in native FP32, the result score was 0.0847633481; the all native FP32 score was 0.0779616311 (a +8.72% shift for this image). Two warm repeats returned the same score. One full prediction took 12.61 s; repeated worker timings were 15.25 s then 12.24 s. The measured process peak was 1.063 GB, below the 2 GB worker budget. These are single-image engineering measurements, not a labeled detection-accuracy evaluation.

The score and noiseprint differences are intentional FP16 quantization effects. Evaluate against the intended workflow before interpreting scores as calibrated probabilities. See validation.json for machine-readable measurements and conversion.json for source checkpoint and converter hashes.

Source and license

TruFor is from GRIP-UNINA, CVPR 2023: Fabrizio Guillaro, Davide Cozzolino, Avneesh Sud, Nicholas Dufour, and Luisa Verdoliva, “TruFor: Leveraging All-Round Clues for Trustworthy Image Forgery Detection and Localization.” The pretrained weights are covered by the same license as the code. A GRIP-UNINA collaborator confirmed in issue #36 that redistributions must retain the original license and remain limited to nonprofit use.

Use, reproduction, modification, and redistribution are restricted to informational and nonprofit purposes. Preserve all original notices and credit the original authors whenever publishing results. The GRIP-UNINA license is in LICENSE.txt; the CMX MIT notice is in LICENSE_CMX.txt. This FP16 conversion does not relicense the source weights. WaifuFinder's separate integration code is not included in this artifact.