Published by WaifuFinder. manifest.json pins 6 model files by SHA-256;
the app verifies every one of them, plus the licence, attribution and validation files,
before installation.
PGTFormer — fused GPU Core ML
Self-contained face-restoration packages for WaifuFinder. Install from Settings → Models → PGTFormer. No Python or conversion is required.
Optimizations
- Both encoder and decoder allow Apple GPU execution.
- The FP32 encoder uses nine native scaled-dot-product attention blocks. Positional embeddings apply to Q/K only; V and attention scaling are preserved.
- The FP16 decoder uses native normalization and prunes neighboring frames only after the final temporal operator.
- Original Conv/BatchNorm ordering is retained. Encoder GroupNorm fusion remains off.
The worker compiles the packages locally and caches them by content and macOS
version. The legacy cpu-codebook-gpu-decoder-v1 package-format identifier is
retained for compatibility; it does not prescribe current device placement.
Validation and measurements
validation.json records eight original-network reference windows, product
preview/export with audio, and alternating trials on an Apple M3 with 16 GB RAM.
The attention-fused encoder reduced encoder time by 13.2%, total inference time
by 6.3%, and peak memory by 5.7% relative to the prior GPU encoder. All eight
output PNGs matched that version byte for byte. Original-network maximum error
was 0.01538, below the unchanged 0.02 limit; mean error was below 0.000150,
against a 0.002 limit. Timings depend on hardware and system load.
Files
PGTEncoder.mlpackage and PGTDecoder.mlpackage contain the models.
manifest.json pins every model file by SHA-256 and records the conversion.
validation.json contains qualification results. LICENSE and NOTICE contain
upstream terms and attribution. WaifuFinder pins the release and verifies all files.
Provenance and licence
Original network: kepengxu/PGTFormer,
revision e78a26cd463182f4a5aa96b8681110da8615d44d.
Original weights: kepeng/pgtformer-base,
revision bc0b7e3ff10cfa0f2f335f5256919de20fe2b29d.
Non-commercial. Upstream is BSD-style with a non-commercial restriction.
Commercial use requires permission from Xidian University or the contributors.
See LICENSE and NOTICE.
- Downloads last month
- 23