Skip to content

v0.5.0-beta.3

Latest
Latest

Choose a tag to compare

View all tags
@github-actions github-actions released this 04 Sep 16:47
· 14 commits to main since this release
v0.5.0-beta.3
0d7ab0c

🚀 XianScan v0.5.0-beta.3

Self-hosted AI-powered manga, manhua, and comic localization engine with embedded OCR, neural text detection, LaMa inpainting, and typesetter.

📦 Multi-Platform Downloads

Platform Architecture File Acceleration Engine
Windows x86_64 xianscan-windows-x86_64.zip ⚡ DirectX 12 DirectML (AMD, Intel, NVIDIA)
Linux x86_64 xianscan-linux-x86_64.tar.gz ⚡ NVIDIA CUDA + cuDNN
macOS Apple Silicon (M1-M4) xianscan-macos-arm64.tar.gz ⚡ Apple Silicon CoreML + Metal

🧩 Browser Importer Extensions

  • Chrome / Edge / Brave / Opera: Download xianscan-importer-v1.2.0.zip and extract to load unpacked in chrome://extensions.
  • Firefox: Download xianscan-importer-firefox-v1.2.0.xpi and install directly in about:addons.

📱 Mihon / Tachiyomi Android Extension

  • 1-Click Extension Repository URL: https://raw.githubusercontent.com/ArbenApura/xianscan-rust/repo/index.min.json
  • Direct APK: Download tachiyomi-all.xianscan-v1.6.2-release.apk below.

What's Changed

  • feat(docker): official container image + GHCR publish in release workflow by @lenaxia in #2

Docker Container Support

  • Added official container images published to GitHub Container Registry (ghcr.io/arbenapura/xianscan) on an Ubuntu 24.04 runtime (contributed by @lenaxia in #2).
  • Includes non-root execution, persistent /config volume support, graceful SIGINT shutdown handling for safe SQLite WAL flushing, and pre-warmed fontconfig caches.

Reasoning Model Support and Token Escalation (Fixes #1)

  • Resolves issue #1 where reasoning models caused silent translation failures, false "done" page statuses, and empty raw prompt records.
  • Added dynamic token budget escalation for reasoning models (GLM-4.7, DeepSeek R1, OpenAI o1/o3) that spend completion tokens on internal thinking steps.
  • Completion token limits automatically double across retries when hitting length cutoffs (up to 65,536 tokens).
  • Full LLM prompt messages are preserved in SQLite telemetry even during request failures, preventing blank [] prompt records.
  • Added explicit error classifications (TOKEN_BUDGET_EXHAUSTED, EMPTY_LLM_RESPONSE, UNPARSEABLE_LLM_OUTPUT) and reader error banners instead of marking failed pages as complete.

Book Translation Directives

  • Added series-specific custom translation directives (custom_prompt schema and BookDirectivesModal) to enforce consistent tone, terminology, and naming rules across chapters.
  • Factored directives into prompt caching fingerprints (SHA-256).

Shrinkwrap Cavity Cleaning for Speech Bubbles

  • Added an outside-in cavity cleaning algorithm in the Rust cleaning pipeline that detects speech bubble contours and protects bubble borders while removing interior text.
  • Added statistical luminance and saturation checks to safeguard translucent and artistic speech bubbles from flat fills.

Targeted Retranslation and Reader Controls

  • Added dedicated single-page translation endpoint (POST /api/pages/[id]/translate) and retry action in the Inspector modal.
  • Added Webtoon continuous reader overlay badges showing active processing steps and inline error retry banners on failed pages.
  • Added multi-page checkbox selection in Grid view with floating action toolbar to retranslate selected or errored pages.
  • Resolved pipeline state reconciliation issues so aborted or failed batch jobs no longer leave pages stuck in loading states.

Prompt Inspector Improvements

  • Redesigned PageLlmPromptModal with dedicated tabs for structured conversation history, token usage telemetry, and raw formatted JSON payloads.

Browser Importer Extension (v1.2.0-beta.3)

  • Added chapter route isolation across URL query parameters and hash changes to prevent stale chapter selections.
  • Migrated to native Blob URLs and direct buffer decoding to avoid UI thread freezes during live imports.
  • Encoded proxy image data as Base64 to prevent [object Object] image corruption when bypassing mixed-content restrictions on HTTPS manga sites.
  • Updated popover tracker with lightweight server thumbnails and lazy image decoding.

New Contributors

Full Changelog: v0.5.0-beta.2...v0.5.0-beta.3

🐳 Docker Image

docker pull ghcr.io/arbenapura/xianscan:0.5.0-beta.3

Run it: docker run -d -p 8124:8124 -v xianscan-config:/config ghcr.io/arbenapura/xianscan:0.5.0-beta.3

All published versions: https://github.com/ArbenApura/xianscan-rust/pkgs/container/xianscan

🛡️ VirusTotal Scan

Artifact Report
release-assets/xianscan-windows-x86_64.zip View report
release-assets/xianscan-linux-x86_64.tar.gz View report
release-assets/xianscan-macos-arm64.tar.gz View report
release-assets/xianscan-importer-v1.2.0.zip View report
release-assets/xianscan-importer-firefox-v1.2.0.xpi View report
release-assets/tachiyomi-all.xianscan-v1.6.2-release.apk View report

Contributors

lenaxia
Assets 8
Loading