v1.0.34

longbridge/gpui-kitv1.0.34Aug 8, 2026by GargantuaX

AI Summary

This release enhances watermark removal capabilities by implementing evidence-gated handling for difficult Gemini 96px watermarks and introduces verification tooling to ensure quality. It aims to improve processing coverage while maintaining content safety.

Key Highlights

Evidence-gated handling for difficult Gemini 96px watermarks at 192px anchors, including low-confidence cases.
Prefers conservative exact-anchor results to avoid damaging nearby text or artwork.
Adds `pnpm issue:verify` command for repeatable issue sample verification with Release asset digest checks.
Revalidation of public samples shows improved detection (14/14) but low strict end-to-end quality passes.

New Features

Evidence-gated watermark handling logic for specific difficult cases.
pnpm issue:verify command for repeatable issue sample verification.

Full Release Notes

## Watermark removal

- Adds evidence-gated handling for difficult Gemini 96px watermarks at the 192px right/bottom anchor, including low-confidence cases that were previously returned unchanged.
- Prefers a conservative exact-anchor result when a shifted candidate would damage nearby text or artwork.
- Keeps visible-residual and uncertainty metadata for imperfect best-effort outputs; variable contour residuals remain under investigation in #120.

## Verification and tooling

- Adds `pnpm issue:verify -- --issue <number>` for repeatable issue sample verification with Release asset digest checks and reviewable reports.
- Tightens diagnostic alpha-profile admission so visibly dirty candidates cannot be promoted based only on relative score improvement.
- Revalidates the 14 public #120 samples with 14/14 watermark detection, 3/14 strict end-to-end quality passes, and no admitted production alpha-profile candidate.

This release improves processing coverage and content safety. It does not close #120 or claim that all visible contour residuals are removed.