Clark Swin2SR ×2 — text-robustness candidate v2 (realistic UI inputs)
A retrained Swin2SR ×2 super-resolution model, continued from the prior
text-robustness candidate with a broader degradation battery and a larger
dense realistic-text corpus. This revision (step-27,500) is the dominant
checkpoint of the v2 run: it improves on the previous ternary release across
all 12 measured realistic degradations.
This is a candidate revision. It is a strict fidelity improvement over the
prior ternary release on a realistic input distribution, but it does not
replace the promoted G3 artifact and does not claim q4 non-inferiority.
What changed vs the previous candidate
- Broader degradation battery (12 total). Added sharpening halos
(unsharp-mask overshoot) and gamma/color shifts to the existing set
(bicubic/lanczos/bilinear/area + 4:2:0 chroma JPEG + WebP + blur + noise +
fractional DPR + clean identity).
- More training data. Added 40k dense realistic-text v2 images (22 fonts,
7 languages, code/UI copy, 8 themes, backdrop-blur panels) on top of the
44k real UI + 23.5k dense-text v1 corpus — 107,515 images total.
- Checkpoint selection. The full v2 continuation ran to step-33,000, but
later steps over-specialized to the new degradations and lost core ground.
Step-27,500 is the dominant point: best mean fidelity with no degraded case.
Measured result (realistic UI distribution)
25 real UI screenshots × 12 degradations; PSNR, edge-PSNR, and OCR legibility
(macOS Vision character accuracy). Paired against the previous ternary release
(g2hard) and the shipped q4 reference.
- Wins PSNR and edge-PSNR on 12/12 degradations vs the previous ternary
release — mean +0.288 dB, minimum +0.125 dB (no degraded case).
- Biggest gains on the newly-trained degradations: gamma +0.61 dB, sharpen
+0.35 dB, blur +0.32 dB, WebP +0.30 dB.
- Dense ONNX is numerically identical to the PyTorch model (133 dB).
- Packed ternary native bundle passes browser-artifact conformance (76 dB).
Honest limits: the candidate still trails q4 on PSNR/edge (q4 is a larger,
higher-precision reference); OCR legibility is roughly on par with q4 and the
prior release (wins 4/12 vs q4). The win is fidelity + size, not OCR dominance.
Artifacts
model.safetensors — dense weights (4.08 MB).onnx/model.onnx / onnx/model.onnx.gz — portable browser graph (input
pixel_values, output reconstruction); gzip 1.68 MB (29.5% of q4).native/swin2sr-textstudy-x2.bdk.gz — packed ternary native bundle (320 KB,
5.6% of q4), conformance-passed.technical/quantization.json, technical/candidate-ledger.json,
technical/browser-conformance.json — full quality receipts and
per-degradation results.
Provenance
Source study: clark-quantize experiments/browser_swin2sr +
realistic_text_eval. Training receipt and evaluation receipts are linked in
technical/candidate-ledger.json.
