arraypress/upscale-swinir-l-real

SwinIR-L real-world ×4 for Core AI

The default upscaler of upscale and imagine --upscale: faithful 4× super-resolution — it recovers detail without inventing it.

• swinir-l-real-x4-128-float16.aimodel: input x [1,3,128,128] float16 in [0,1] (RGB), output output [1,3,512,512] in [0,1]. Tile any picture with ~16 px overlap and feather the seams (swift-upscale does).

Why this one

Chosen in a bake-off of seven upscalers (×4 shrink-and-back on 5 generated and 5 stock photos, LPIPS lower is better): SwinIR-L scored 0.185 / 0.208 (generated / real) and 0.213 / 0.282 with blur, noise and JPEG damage, beating Real-ESRGAN x4plus (0.195 / 0.236) in every cell, with the best PSNR (28.05 / 24.45 dB). It keeps hair and foliage natural where Real-ESRGAN smooths them.

Measured (M3 Max, macOS 27, GPU)

• Parity with the PyTorch float32 model, per 128-px tile: 66–68 dB PSNR (float16; float32 export 124–125 dB). Float16 is shipped: half the size, slightly faster, no visible difference.
• Tile size: 128 px had the lowest cost per new pixel — 24.9 µs, against 26.9 at 256 px and 29.0 at 192 px.
• Seams: a 256-px picture upscaled as overlapping tiles against PyTorch on the whole picture in one run: 41.6 dB overall, and the seam bands match as well as everywhere else (41.8 dB) — no seams.
• Speed: 1024×1024 → 4096×4096 in 25–29 s (three runs on a quiet machine; first load 0.2 s once compiled) — about 0.25 s per 128-px tile, so time grows with the number of tiles.

Provenance

• Weights: SwinIR v0.0 release, 003_realSR_BSRGAN_DFOWMFC_s64w8_SwinIR-L_x4_GAN.pth, byte-identical to the release asset (SHA-256 99adfa91…328c0a).
• Export: export_swinir.py (this repository) — spandrel 0.4.1 loads the checkpoint; the attention masks are precomputed for the 128-px tile (SwinIR rebuilds them per call at other sizes, which crashed the converter); coreai-torch 0.4.3, coreai-core 1.0.0b3, float16 under autocast.
• Licence: Apache-2.0 (LICENSE), with NOTICE.