Repository navigation
PaperReader v2.1.12
📖 Read the User Guide (中文) before downloading — installation, first-run setup, and usage. 中文发布说明
PaperReader v2.1.12 preserves PDF text selections when opening the context menu and crops LaTeX figure and table thumbnails to the actual artwork.
What's new
- Stable PDF selections. Opening the context menu keeps the selected text visible in both PDF panes and in scroll/single-page modes. Right-click and macOS Control-click retain the selection for counterpart lookup, Ask AI, and annotations.
- Tight LaTeX thumbnails. Each Figure/Table preview is cropped from the compiled PDF around the artwork beside its caption — vector graphics, embedded images, multi-panel composites, and booktabs-style rules — for both the original and the translated PDF, each located independently. The previous float-anchor fallback remains for pages without usable artwork objects.
- Smarter caption lookup. Captions that start their own line win over mid-sentence references such as "in Table 2.". When text extraction merges subfigure labels into the caption line ("…bFigure 2:…"), the caption is still found and cropped correctly.
- No pipeline change. LaTeX projects keep their existing single compile per side; no extra MinerU call is needed for thumbnails.
Compatibility
Existing accounts, annotations, provider settings, documents, and translation checkpoints remain usable. Keep the same DATA_DIR and AUTH_SECRET_KEY. No manual database migration is required. Stored thumbnails refresh automatically when a compiled PDF changes.
Verification
- Chromium and WebKit browser checks pass for text selection, context menus, Ask AI, counterpart lookup, and annotation notes in both PDF panes and reading modes, including macOS Control-click. The frontend production build passes.
- New backend tests cover artwork cropping beside figure and table captions on both PDF sides, caption-line preference over body references, and the merged-subfigure-caption fallback. The full backend suite and
compileallpass. - The arXiv
2605.18309LaTeX archive compiles with its declared engine; all 6 Figures and 2 Tables receive tight thumbnails on their correct pages (2, 6, 7, 8, 19, 19, 20, 21) for both PDFs. - The release workflow tests the backend across Windows, macOS, and Linux, builds both desktop applications, and smoke-tests the packaged executables before publication.
Limitations
Artwork detection relies on embedded vector/image objects and horizontal table rules. Floats without such objects (for example, plain tabular arrays without rules) fall back to the previous anchor-based crop or a page preview. The local regression does not call external LLM or MinerU services.
Downloads
- Windows x64:
PaperReader-v2.1.12-Windows-x64.zip - macOS Apple Silicon:
PaperReader-v2.1.12-macOS-arm64.dmg - Frontend/API version:
2.1.12; Git tag:v2.1.12.