SwinIR-L real-world super-resolution x4 (GAN), converted to Core AI.

Original model and weights: SwinIR, Jingyun Liang et al. (https://github.com/JingyunLiang/SwinIR),
Apache License 2.0 (see LICENSE). Checkpoint: 003_realSR_BSRGAN_DFOWMFC_s64w8_SwinIR-L_x4_GAN.pth
from the v0.0 release (SHA-256 99adfa91350a84c99e946c1eb3d8fce34bc28f57d807b09dc8fe40a316328c0a).

Modifications: exported to Core AI (float16, fixed 128x128 -> 512x512 tile) with coreai-torch 0.4.3 /
coreai-core 1.0.0b3; attention masks precomputed for the 128-px tile; output clamped to [0, 1].
The export script is export_swinir.py in this repository.