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.
SAFIRE native Swift / MLX
Official SAFIRE pretrained checkpoint converted without quantization for WaifuFinder’s Photos-only Image Lineage workspace.
- Original work: Myung-Joon Kwon, Wonjun Lee, Seung-Hun Nam, Minji Son, Changick Kim, “SAFIRE: Segment Any Forged Image Region,” AAAI 2025.
- Upstream: https://github.com/mjkwon2021/SAFIRE at ec2af6af8c789fb6359493a3e352f89c6fd8107f.
- Weights: CC BY 4.0; code architecture: Apache 2.0; independently authored WaifuFinder code: MIT. Retain all notices when redistributing.
- Source checkpoint SHA-256: 65bdf6880380162d57f77d71b3fd367c3f16aebca316f942a821da16e9225b12.
The 348 FP32 parameter tensors preserve upstream storage bytes. Swift prepares convolution layouts once. Runtime requires Apple Silicon, macOS 15+, mlx-swift 0.31.2 / MLX core 0.31.1. No Python, PyTorch or OpenCV is required for production inference.
Supported outputs are automatic binary manipulation probabilities and multi-source region partitions, using all 256 grid prompts. Cosine clustering retains raw-feature mean centers, with deterministic farthest-first initialization for reproducibility; upstream’s default initializer is random.
Native acceleration includes a fused fixed-point RGBA resize/normalize Metal kernel, GPU FFT high-pass filtering, chunked relative-position SDPA, exact GELU, fused MLX LayerNorm, cached compiled two-way decoder graphs, and warm worker reuse. Full precision is retained.
validation.json records encoder intermediates, decoder masks, automatic probabilities, partition agreement, preprocessing and orientation checks on two synthetic diagnostic images. These checks establish numerical/spatial parity for those inputs, not real-photo forensic accuracy. Probability threshold 0.5 is a display convention, not proof of manipulation. profiles.json contains actual M3 screening times under varying background load. Independent M4 validation and paired full-inference measurements select a one-prompt decoder batch; M3 retains batch four. Final complete warm-worker medians are about 4.3 s (M3) and 3.1 s (M4) for this diagnostic image. These are not whole-Photos-family throughput claims.
SAFIRE does not identify an editing application, reconstruct chronology, or establish AI restoration history. A source partition is not a semantic object label. Use pixel correspondence and other independent evidence, and keep unresolved histories uncertain.
Quantized