Skip to content

v0.6.4 | 450 Formats, Faster Previews & Search

Choose a tag to compare

@orca-zhang orca-zhang released this 05 Nov 04:40
· 30 commits to main since this release
Download Link Download size
CLI Lite: darwin-arm64 filens-0.5.6-macos-arm64-lite.tar.gz 19.08 MiB
CLI Full: darwin-arm64 filens-0.5.6-macos-arm64-full.tar.gz 33.21 MiB
CLI Lite: linux-x64 filens-0.5.6-linux-x64-lite.tar.gz 18.92 MiB
CLI Full: linux-x64 filens-0.5.6-linux-x64-full.tar.gz 34.87 MiB
CLI Lite: linux-arm64 filens-0.5.6-linux-arm64-lite.tar.gz 18.81 MiB
CLI Full: linux-arm64 filens-0.5.6-linux-arm64-full.tar.gz 32.56 MiB
CLI Lite: win32-x64 filens-0.5.6-windows-x64-lite.zip 18.86 MiB
CLI Full: win32-x64 filens-0.5.6-windows-x64-full.zip 35.09 MiB
Browser runtime: split filens-preview-0.3.4-split.zip 6.36 MiB
Browser runtime: all-in-one filens-preview-0.3.4-all-in-one.zip 11.90 MiB
iOS source SDK filens-ios-sdk-0.1.4.zip 6.38 MiB
Android source SDK filens-android-sdk-0.1.4.zip 6.39 MiB
dsh-filens 0.1.6 dsh-filens-0.1.6.tgz 6.53 MiB
Checksums SHA256SUMS SHA-256
Build manifest filens-release-v0.6.4.json Versions, sizes and validation

13 package downloads + manifest + checksums (15 assets). Full CLI includes OCR; Lite keeps OCR optional. Native bundles include the offline viewer. Mobile downloads are source SDKs. MiB = 1,048,576 bytes.

CLI installers and library modules are available from npm: filens-cli, filens-cli-lite, filens-preview, dsh-filens and @filens/*. Platform/OCR dependencies install automatically through npm; use the Full/Lite archives above for offline installation.

What's new

Compared with v0.6.3, FiLens expands browser previews from 402 to 450 extensions and CLI inputs from 334 to 349 formats, reduces format-specific startup work, and improves repeated document previews and local search. Parsing and conversion remain in the shared frontend/WASM core.

Components: CLI 0.5.6, browser preview 0.3.4, iOS/Android source SDKs 0.1.4, DSH plugin 0.1.6, and updated @filens/* modules. The CLI and browser counts describe different capabilities; inspect filens capacity --full --json for each CLI operation.

DSH 0.1.6 update — November 5, 2026

  • Keep zoom out / percentage / zoom in / fullscreen beside the DSH address bar for every format. Remove page, rotation and manual fit buttons and warning-count badges; retain PPTX keyboard navigation.
  • Reuse native zoom and provide shared view zoom/fullscreen for text, Word, archives and media. Word auto-fit follows pane width independently of user zoom.
  • Keep CAD layers and the 3D inspector as compact entries. Group worksheet, archive and attachment controls under Tools, retaining their actual nodes and handlers. Keep audio/video playback and progress below the player, and size the zoom percentage to four characters. Controls move into fullscreen and return to the address bar on exit.
  • Validate 69 Vue tests, 18 DSH tests, 16 isolated format-loading paths, 12 complete browser previews and six actual Desktop formats. Browser fullscreen passes; native Desktop fullscreen is excluded.
  • Shorten the English/Chinese README DSH section and expand the featured comparisons to 12 measured scenarios across DSH, XLSX, ZIP, PPTX, RAW and STEP, with measurement dates and stage boundaries. Compact-toolbar checks.

DSH 0.1.5 update — November 5, 2026

  • Rename the preview entry to FiLens Document Preview and simplify its description to local previews for 400+ formats. The separate FiLens Tools entry retains its reading, OCR, search, excerpt and conversion description.
  • Move format controls beside the address bar: page navigation, zoom, rotation, fullscreen, worksheet tabs, archive search/sort, Office attachments, audio/video playback and the 3D inspector. Remove manual Fit / Width / Height buttons in both the header and fullscreen.
  • Keep actual control nodes, handlers and independent tab state. Hide retained parent controls while an attachment is open. Move controls into the fullscreen preview and return them to the header on exit.
  • Fix narrow-header clipping, nested scroll containers, blocked button clicks and the misplaced 3D tools opener. The inspector opens below the header without shrinking the canvas, with keyboard tabs, Escape and focus restoration.
  • Re-record the 3D / panorama GIF, WebP and MP4, including header model/view/section tools, panorama zoom and video seeking/playback. Update the FiLens skill, README SOTA shortlist and measured bundle/package sizes.
  • Validate 66 Vue tests, 17 core attachment tests, 18 DSH tests, 16 isolated format-loading paths, 12 complete browser previews and six actual Desktop formats. Browser fullscreen passes; the Desktop run explicitly skips native fullscreen. 0.1.5 checks and exact package bytes.

DSH 0.1.6 uses CLI 0.5.6 and embeds the shared-source preview runtime 0.3.4 with the header-control changes. The standalone browser/npm and SDK archives keep their existing payloads. The sidebar timing table below retains its measured 0.1.3 snapshot; the 0.1.6 interface update does not introduce a new latency claim.

Formats and features

  • Add filens inspect FILE --json to identify actual file content independently of the filename, with MIME, evidence categories and extension-mismatch reporting. The shared WASM detector reads at most 256 KiB; large inputs are not copied into the temporary workspace. Ambiguous text/containers and unknown binary remain explicit. Detection evidence does not imply full-file validation. Refresh the English/Chinese terminal GIFs with an actual renamed-file inspection.

  • CLI reading, OCR, excerpts and directory/archive search now accept TGA, DDS, DIB, CUR, PCX, JPEG XL, QOI, Farbfeld, PSD, PSB, ICO, ICNS, AI, EPS and PS through the existing shared WASM decoders. Photoshop uses saved composites; AI uses PDF-compatible artwork or the PostScript interpreter; raster-only artwork requires OCR for text. Existing image and design conversions are also available in the CLI build.

  • Add TIFF, TGA, PBM/PGM/PNM, DDS, HDR, DIB, CUR, PCX, EXR, JPEG XL, QOI and Farbfeld previews, plus DWG-to-DXF 2D preview. The optional graphics engine loads on first use. TIFF displays the first usable image; HDR/EXR previews use 8-bit output. See the support matrix.

  • Expand archive browsing and nested opening to BZip2, XZ, LZMA, Zstandard, Unix compress, LZMA86, compound TAR extensions, AR/DEB, CPIO, CAB, ISO/UDF, RPM, XAR, WIM/ESD and LHA/LZH, including WIM multi-image directories and UDF images named .iso.

  • Add XPI, CRX2/CRX3, ZIPX with BZip2/Deflate64/LZMA entries, and CBR/CB7/CBT containers, using actual readers for their contents.

  • Archive browsing gains name filtering, name/size/date sorting, empty-state messages, arrow-key navigation, Enter/Space to open, and Escape to clear filtering. Returning from a preview restores the query, scroll position and keyboard focus.

  • Add the optional @filens/converter-client/format-detector export. Importing it does not fetch WASM; instance creation loads the selected SIMD/scalar artifact. SIMD download, compilation or initialization failure falls back to the scalar build.

  • Expand image, code/text, archive and system-package examples. Remote opening handles encoded URLs, signed query strings, explicit filenames for extensionless URLs, same-origin relative paths, and browser history.

  • DSH mounts previews directly in the sidebar DOM. Version 0.1.6 uses a compact address-bar toolbar and dedicated tool menus; the 3D/panorama gallery records 0.1.5; the installation demo and timed 0.1.3 snapshots retain their own provenance.

Performance improvements

Loading, compression and package size

  • Load preview drivers by format, separating routing metadata from renderers. Share pending and successful imports. Text/PDF no longer preload Office, iWork, CAD or Three.js; nested files and conversion outputs follow the same routes.
  • In the DSH loading comparison, entry JavaScript fell from 3.04 MiB to 345 KiB, about 89%; TXT first-open runtime requests fell from 3.09 MiB to 449 KiB. These are requested raw runtime bytes, excluding the input file. Measurements.
  • Load the DOCX engine, JSZip compatibility reader, encrypted-ZIP reader and OLE attachment reader only when needed.
  • Load GLB/glTF, OBJ/MTL, USDZ, LDraw, FBX, 3DS and WRL/VRML engines for their respective formats. FBX, 3DS and VRML have separate chunks; public synchronous loader APIs remain available. The all-in-one build still embeds every engine. Media/readers, FBX/3DS, VRML.
  • Reuse fflate for JSZip DEFLATE encoding and Pako for ZIP/TIFF decoding, removing duplicate production codecs while retaining metadata, CRC and decompression errors. Compression.
  • Pack shared WASM, workers, CSS and large static data into the all-in-one runtime, with supplied gzip/Brotli representations of the main JS verified at no more than 4,000,000 bytes each. Archive sidecars, HTML, notices and the complete npm package are measured separately. Initialization includes decoding work. Delivery and startup tradeoffs.
  • Exclude source maps from published frontend packages and remove invalid declaration-map references. Local builds retain hidden maps. This reduces package download/extraction bytes rather than claiming equivalent runtime-JS savings. XLSX, core/archive, other packages.

ZIP, XML, media and resources

  • Share ZIP directory indexes, lazy decompression and bounded entry caches across EPUB/CBZ, XMind/MindManager, Visio and 3MF. Retain ZIP64 and legacy Unicode compatibility; an unused attachment's unusual compression no longer forces a complete PPTX ZIP rescan.
  • Inflate large ZIP entries directly into the destination buffer, reducing intermediate chunks and concatenation. Ordinary resource paths bypass repeated normalization; actual dot segments retain normalization. ZIP reader.
  • Use the fast OpenXML parser for ordinary PPTX; load the general XML parser for syntax such as special entities. Load it after recognizing XLSX, and skip it for CSV. Remove unused XMLBuilder code from production previews and reduce temporary PPTX node arrays. XML, shared readers.
  • Avoid an extra full byte copy when creating media Blob URLs, and share the preview/download URL. Shared buffers still use snapshots. Prefer native byte-to-Base64 APIs with bounded fallback chunks across media, EPUB and 3D textures.
  • Stop TIFF preview decoding after the first valid image with usable dimensions and pixels, avoiding work on undisplayed pages. Full-conversion validation remains separate. TIFF.
  • Read only referenced EPUB resources and scan CSS URL/import candidate tokens. Find direct MindManager topics in wrapper order without repeated descendant walks; skip unnecessary attributes on invisible Visio containers while still parsing child shapes. Document resources.
  • Reuse normalized uploaded-resource paths, root aliases and basename candidates, preserving exact-path priority and duplicate-name ambiguity. Build KMZ case-insensitive and unique-suffix indexes only when a fallback is needed. Resource indexes, KMZ.
  • Share extension/MIME indexes for default format routes, preserving first-rule and extension priority. Custom rule arrays remain live; callers can explicitly create a snapshot index.
  • Reuse archive name ordering, recompute size/date ordering only when options change, and filter sorted results. Keyboard movement uses adjacent result rows instead of collecting the entire DOM list.

XLSX parsing, formulas and reopening

  • Build cell models directly for numeric/shared-string sheets, including supported column styles and simple sheetData attributes. Cache styles for accessed columns and move row defaults outside cell loops. Complex XML, rich text and formulas retain the general-parser fallback. Columns, XML fast path.
  • Reuse attribute/numeric matchers, retain only used cell attributes, and reduce repeated property wrappers, column-name conversions and range-bound parsing.
  • Resolve formula dependencies by row/column coordinates. Build bitmap indexes directly for ordered sheets; sort a private index for unordered inputs, retaining last-declaration semantics for duplicate coordinates. Cell indexes, compact indexes.
  • Evaluate ranges over existing cells and use contiguous data for complete rows. Stream SUM accumulation without temporary range arrays or a second pass, retaining argument order, overflow, circular-reference and error behavior. Streaming formulas.
  • Reuse recently parsed workbook models to skip repeated ZIP/XML parsing. Tabs receive independent cells/fonts; formula evaluation and font mapping cannot contaminate subsequent opens. CSV keeps direct parsing, and public parser calls still parse each input.

PDF first page, scrolling and reopening

  • Reveal a document when its first page has actually painted; render nearby pages in the background. Nonvirtual print/export still waits for complete rendering. DSH removes the PDF loading overlay in the same frame.
  • Locate visible pages with ordered page slots and binary searches over actual DOM geometry. Cancellation/release visits only pages with active canvases/tasks. Viewport lookup.
  • Reuse an independent snapshot of the complete first-page pixels for recent reopens, keyed by full-file SHA-256, exact CSS viewport and canvas resolution: 4 pages, 16 MiB, 30-second idle expiry. Password, URL and nonvirtual print paths bypass the cache. Every open still loads/validates the PDF; closing releases documents, workers and active canvases. Implementation and checks.

iWork, fonts, CAD, SVG and 3D

  • Index iWork text styles and Pages paragraph/inline-attachment positions with ordered ranges, preserving duplicate positions, empty-style inheritance/reset and malformed-position fallbacks. Text indexes.
  • Reuse normalized font names and glyph state while collecting fonts; stop accumulating after the sample limit. Font diagnostics reuse classification within one call but return independent objects and recompute across calls. Font pipeline.
  • Maintain a document-local CAD layer-name set for bulk/single-layer selection, synchronized on replacement, cancellation, failure, redraw and destruction.
  • Share delegated SVG events and ID indexes for mind maps/diagrams; selection updates only previously/newly selected nodes. Preserve keyboard navigation and duplicate-ID highlighting; detach old handlers on page changes/destruction. SVG selection.
  • Share texture lookup and scene-finalization work across five 3D loaders, preserving resolution priority, warnings, camera state and object-URL cleanup.
  • Reuse parsed Pages/Keynote models without starting a parsing worker on cache hits. Numbers still starts with the first sheet and restores archive context only when another sheet is requested after reopening.
  • Reuse CPU meshes from STEP's WASM conversion while recreating independent geometry, materials, scenes and GPU resources per tab. Direct public loadStepObject calls default to conversion; previews explicitly enable short-term reuse.
  • Share a 4-entry, 64 MiB estimated model budget, 30-second idle-expiry LRU across XLSX/iWork/STEP. Retain data snapshots only; closing still releases DOM, URLs, workers and GPU resources. Password-based opens, including decrypted Office outputs, bypass reuse. Full-content digests invalidate changed files. Model cache and costs.

CLI search, OCR and content detection

  • Determine exact-search body matches using the current file's result boundary instead of rescanning previous files' blocks; skip tokenization unused by exact-mode scoring.
  • Accumulate BM25 document frequency by query-term ID and calculate IDF/query weights once. Reuse normalized body/filename phrase checks and use a lowercase fast path for ASCII.
  • Store query features in compact ordered numeric lists and extract filename frequencies on demand. Count body block lengths/query frequencies directly instead of building complete dictionaries for unrelated words. Borrow normalized UTF-8 slices to reduce temporary word/Chinese-bigram allocations.
  • Select the best result files before sorting their candidate excerpts, preserving matched-file totals, stable ranking, up to three nonoverlapping excerpts and source offsets.
  • Add SIMD ASCII detection, lowercasing and separator scanning to the shared search core: optional SIMD128 in browsers, SSE2/NEON in native CLI, scalar fallback elsewhere. Mixed Unicode retains the NFKC path; execution errors do not silently retry operations.
  • Index OCR score records by line after one scan, Office media by slide path/first reference, and page selections with ordered lookup. Retain output order and score semantics.
  • Preserve Chinese/Unicode tokenization, numeric-token thresholds, BM25 floating-point operation order, ranking, excerpts and source mapping. Individual benchmark-round speedups are not multiplied into a cumulative claim. Round 1/OCR/routing, exact search/media, query calculations, compact features, filename frequencies, selection, direct extraction, borrowed slices, SIMD.
  • Add SIMD128 text scanning to content detection and reuse validated ASCII prefixes before later Unicode validation, preserving BOM/control-character/truncated-UTF-8 decision order. Detection measurements.

Bug fixes and compatibility

  • RAR/7z after production bundling: copy archive WASM/worker resources when a Vite consumer rebundles the all-in-one npm export, preserving reachable resource URLs.
  • Large SUM ranges: remove function-argument spreading; a 160,000-cell range works while preserving floating-point order, intermediate overflow and error short-circuiting.
  • Unicode EPUB styles: encode strings as UTF-8 before Base64 inside opaque-origin frames, avoiding direct btoa on Unicode strings.
  • HEIF colors and auxiliary images: fix 10-bit Apple-edited HEIF and grid/nested-grid color-metadata inheritance; follow image-tile relationships without treating alpha, thumbnails or tone maps as RGB configuration. Add P3, linear-ICC and alpha regressions.
  • 10-bit HEVC 4:4:4: update chroma decoding and verify reference pixels.
  • Unused ZIP attachments: defer unsupported-compression errors until the affected attachment is actually read, allowing supported document parts to open.
  • ZIP sizes and cancellation: reject mismatches between declared size and actual decompressed output; check cancellation even on cached reads.
  • Malformed images and optional engines: validate decode bounds before dimension access; reject truncated PCX rows and invalid CUR directories; report the required optional graphics module explicitly.
  • TIFF later-page corruption: a corrupt undisplayed page no longer prevents previewing an already decoded valid first image; full conversion keeps its validation.
  • Remote loading/workbench state: clear stale previews after download failure while retaining the attempted source; ignore late remote responses after local selection; restore matching sources through browser history. Late route imports no longer override manual category selection.
  • Stale asynchronous results: strengthen cancellation/generation checks for reading, dynamic imports, parser workers and rendering; release old scenes, textures, URLs and events on replacement/destruction.
  • Archive focus/events: fix focus, scroll and empty states after filtering/sorting/returning; detached lists, SVGs and controls cannot mutate the active document.
  • Published type imports: retain the @filens/media package reference in Vue declarations instead of repository sibling paths; remove broken source-map references from published/copied declarations.
  • Direct DSH mounting: handle Desktop cross-origin modules/WASM, Blob workers and shared-style lifetimes. Pass copied source bytes so worker transfer cannot damage the host's retained file; isolate tab controls and destruction.
  • Cache correctness: prevent cross-tab font/formula contamination, stale reuse after same-name edits, wrong PDF resolution after viewport changes, missing Numbers archive context on later sheets, and shared STEP geometry/material destruction. Cache failures return to normal parsing/rendering.

Selected performance results

The latest FiLens implementation is compared with named baselines. Ratios use baseline median ÷ FiLens median. The larger parser table selects substantial inputs and speedups; the complete raw reports retain all controls. The DSH table also includes the PDF comparison.

Complete sidebar previews

Sample · first / cached median (ms) FiLens 0.1.3 DSH built in First-open speedup
PPTX 135 / 81 1830 / 124 13.56×
DOCX 111 / 94 1827 / 112 16.46×
PDF 190 / 77 329 / 110 1.73×

Nine paired rounds per format on Apple M4 Pro/macOS arm64 and DSH Desktop 0.2.1-rc.5, using 659,484-byte PPTX and 601,507-byte DOCX samples. Ratios describe first opens; first means first for that file/process, excluding application startup and retaining system/persistent caches. DOCX follows PPTX. The built-in Office path converts to PDF; FiLens renders directly. All samples, ranges and method.

PDF uses a separate nine-round paired run on the same machine after first-page pixel reuse, with a 689,455-byte sample. Cached opens are 1.43× faster than built in (110 → 77 ms, 30.0% less time). First medians use nine opens and cached medians pool the next 27; each round starts a fresh process per renderer. Timing includes document reading, validation and rendering to ready plus two animation frames, excluding application startup and retaining OS/persistent caches. Each open still validates the PDF and closing releases its worker. PDF samples, ranges and method.

Larger parsing workloads

Larger parsing workload Baseline Baseline / FiLens median (ms) Speedup
XLSX · 2,000 × 64, 128,000 cells SheetJS CE 0.20.3 150.94 / 61.72 2.45×
XLSX · 8,000 × 64, 512,000 cells SheetJS CE 0.20.3 615.71 / 213.50 2.88×
ZIP · 10,000 entries, 18.34 MiB expanded JSZip 3.10.1 700.55 / 55.52 12.62×
ZIP · 2,500 entries, 36.23 MiB expanded JSZip 3.10.1 247.99 / 95.37 2.60×

These measure complete parser calls on identical synthetic inputs, not UI previews. Two warmups and nine measured rounds alternate the libraries. All cell values or complete entry text match; imports, generation, verification, GC and rendering are outside timing. Feature/fidelity scopes differ. The summary selects at least 100,000 cells or 16 MiB expanded content, baseline latency of at least 100 ms and speed ratios of at least 2×. No overall state-of-the-art ranking is claimed. Methods, library versions, raw samples and reproduction.

PDF first-page pixel reuse and XLSX/iWork/STEP model reuse retain their separate nine-round measurements and costs in the PDF report and model-cache report. PDF still validates each open and releases workers. The model-cache comparison observed a 34 ms (17.6%) Numbers first-open regression; small XLSX had no stable gain. These controls remain in the raw data. The 64 MiB model budget is estimated accounting, not a hard JS-heap/GPU cap.

The installation GIF/WebP/MP4 uses public 0.1.3 with CLI 0.5.6. Installation demos and historical measurements.

Validation and limits

  • Full CLI bundles include OCR runtime/models; Lite keeps OCR optional. Both include offline preview assets. Mobile ZIPs are source SDKs, not precompiled AAR/XCFramework binaries.
  • DSH 0.1.6 depends on CLI 0.5.6 and embeds the shared preview runtime 0.3.4. CLI/OCR installation bytes are separate from the plugin tarball.
  • See release validation, exact artifact sizes, format-loading checks, PDF checks and model-cache checks. Each record identifies its own build/measurement stage.
  • Model reuse passed core/XLSX/iWork/3D/PDF/Vue/resource tests, 15 isolated browser format-loading paths, 12 complete preview types and six Desktop preview/zoom/3D-interaction cases. Desktop native fullscreen-exit automation hung and was explicitly skipped on rerun; browser fullscreen checks passed.
  • macOS arm64 native execution is verified; Linux x64/arm64 and Windows x64 are cross-built with the checks recorded in the validation report. No new physical-mobile-device validation is claimed. Fidelity, fonts, browser codecs, WebGL, DWG and HDR/EXR color limits follow the support matrix.
  • Requested runtime bytes, compressed files, npm package sizes, parser microbenchmarks and complete sidebar timings use separate measurements. Percentages from successive optimization rounds are not added together.

First-party code is MIT; retain bundled third-party notices. SHA-256 checksums accompany the release assets.

Use FiLens on ZimaOS

ZimaOS: a personal cloud with FiLens integration

Bring local document previews to your personal cloud. Install FiLens from the official ZimaOS Mod Store, without Node.js or npm:

sudo zpkg update
sudo zpkg install filens

Upgrade an existing installation with sudo zpkg upgrade filens. Download exported offline HTML to your computer to view it; ZimaOS does not launch a browser on the server. Installation, preview export and OCR configuration.

ZimaOS ModStore uses its own component versions: 0.5.1 (CLI 0.5.0, preview 0.2.0).