alastorid/TruFor-Swift-MLX

Published by WaifuFinder. manifest.json pins 2 model files by SHA-256; the app verifies every one of them, plus the licence, attribution and validation files, before installation.

TruFor native weights for WaifuFinder (Swift/MLX)

Converted FP32 weights of the official GRIP-UNINA TruFor checkpoint (CVPR 2023) for the native Swift/MLX runtime bundled with WaifuFinder. The conversion extracts the 927 inference tensors of the official TruFor.pth archive without importing the upstream Python stack.

• Source checkpoint SHA-256: ac1d90e329a72e0d66e8665e123a19e94bfae3209c3ef8a4f9ca3b91578c7844
• Converted weights SHA-256: pinned in manifest.json and re-verified on install
• Preprocessing: native oriented RGB, uint8 / 256, no resize and no padding
• Detection: official confidence-weighted statistics plus the trained 8-128-1 MLP
• Licence: GRIP-UNINA informational/nonprofit only; CMX components MIT. Full terms in LICENSE.txt and LICENSE_CMX.txt. Credit GRIP-UNINA whenever results obtained with this software are made public. On 2026-10-25 a TruFor collaborator confirmed in grip-unina/TruFor#36 that the pretrained weights are covered by the same licence as the code, which restricts use to nonprofit purposes, and that all redistributions retain the original licence and remain subject to its terms, including the nonprofit-use restriction.
• Validation: validation.json records native-vs-reference parity for the detection score and the genuine anomaly, reliability and Noiseprint++ maps at native odd resolution (mask IoU 1.0, score error within 1e-5).