Onyx-Real-ESRGAN

OnyxModels — Real-ESRGAN x4 upscaler (Onyx MLX build)

The learned AI image upscaler for Onyx (the "Enhance (AI)" tier). A ×4 single-pass super-resolution GAN (RRDBNet) that synthesizes real detail — unlike the Lanczos "Quick" tier, which only interpolates.

Layout

mlx/
  model.safetensors     Onyx MLX build — fp16, 33.4 MB (what the app downloads from HF)
  onyx_mlx.json         arch config (rrdbnet, scale 4, num_feat 64, num_block 23, num_grow_ch 32)
pth-src/
  RealESRGAN_x4plus.pth original fp32 PyTorch weights (~64 MB, provenance / rebuild source)
convert.py              reproducible converter (.pth → Onyx MLX safetensors)
verify/
  input_768x512.png     test input
  onyx_esrgan_4x.png     Onyx MLX output (3072×2048)
  pytorch_ref_4x.png     PyTorch reference output — Onyx matches it (99.8% of pixels within ±3, mean diff 0.09)

Provenance

Source: RealESRGAN_x4plus.pth from xinntao/Real-ESRGAN v0.1.0 (Apache-2.0 — commercially shippable). The Onyx MLX build is a straight format conversion (convert.py): keys kept 1:1 with the PyTorch names, conv weights transposed to MLXNN's (C_out, kH, kW, C_in) layout (with np.ascontiguousarray so the buffer order matches the shape — a view-vs-copy bug here scrambles the weights), fp16 storage. No retraining.

Architecture (RRDBNet): conv_first → 23× RRDB (each = 3 ResidualDenseBlocks, dense 5-conv, growth 32) → conv_body + global residual → nearest-upsample ×2, conv_up1 → nearest-upsample ×2, conv_up2 → conv_hr → conv_last. LeakyReLU(0.2). Runs in fp32 compute (the fp16 export is upcast at load; 23 dense blocks overflow fp16 to NaN otherwise). Swift model: OnyxDiffusion/Families/Upscale/RRDBNet.swift + ImageUpscaleModel.swift (tiled, overlap-blended forward for memory-safe large images).

Verified

Device-parity: the Onyx MLX build matches the PyTorch reference within fp16 rounding (max abs diff 0.09 mean, 99.8% of pixels ≤3). Runs on Mac in ~4.5 s for 768×512 → 3072×2048 (single-shot; tiled for larger).

Publishing to Hugging Face

App downloads from wabibito/Onyx-Real-ESRGAN (mirror convention, like the other onyx-* image models). Upload needs a transient HF token you provide and revoke after. The pth-src/ original stays local-only.