Changelog

Added

• Make an animation (GIF maker). Turn 2–1000 still images (JPG, PNG, WebP, HEIC…) into one animated GIF, animated WebP or MP4.
  • Web UI: a "Make an animation" mode with a frame strip. Reorder frames by dragging or with ← →, then pick the format, time per frame, loop count, size, fit (borders or crop), border colour and quality. Redo re-makes it with new settings.
  • CLI: compress-media animate [--format gif|webp|mp4] [--delay ms | --fps n] [--loop n] [--max-dim px] [--fit contain|cover] [--background hex] [--quality 1-100] [-o file]. Folders are sorted by name with numbers in order (shot-2 before shot-10).
  • HTTP API: POST /api/animations, taking multipart frames or chunked upload ids. GET /api/config reports animationMaxFrames.
• Subtitles. Turn the speech in a video or audio file into subtitles with times and text (SRT or WebVTT), ready to upload to YouTube, using whisper.cpp on your own server. Or put subtitles into the video.
  • Speech recognition in about 100 languages, including Vietnamese: the language is detected automatically or chosen, and you can translate the subtitles into English. There are five speech models, from tiny (75 MB) to large-v3-turbo (547 MB); each is downloaded once, on first use. Silence and music are skipped.
  • Subtitles are split into readable cues of at most two lines.
  • Add subtitles to a video as a track viewers can turn on (no re-encode), or burn them into the picture for TikTok, Reels and other apps that ignore subtitle tracks.
  • Use your own .srt/.vtt instead of speech recognition.
  • Web UI: a "Subtitles" mode. Drop a video, or a video plus its .srt with the same name. View & edit shows the video with its subtitles next to an editor; Save makes the result again from your text. Redo switches between SRT, VTT, track and burn-in without listening again.
  • CLI: compress-media subtitles [--lang vi] [--translate] [--model small] [--format srt|vtt] [--embed none|track|burn] [--font-size …] [--srt file].
  • HTTP API: POST /api/subtitles and GET /api/jobs/:id/subtitles?format=srt|vtt. Jobs report a stage (model, transcribe, embed).
  • Configuration: WHISPER_MODEL, WHISPER_MODELS_DIR, WHISPER_DOWNLOAD (off for servers without internet access), WHISPER_MODEL_URL, WHISPER_PATH, WHISPER_THREADS.

Changed

• The mode switch (Compress files, Make an animation, Subtitles) sits at the top of the page, above the settings and the file list.
• The Docker image includes whisper.cpp and the DejaVu fonts (about 100 MB more; --build-arg WHISPER=false leaves whisper.cpp out). Speech models are kept in /data/models, on the data volume.

Fixed

• Animated GIF or WebP input is no longer reduced to its first frame when the chosen image format can't animate. AVIF output becomes animated WebP, and JPEG/PNG output stays GIF. The result says so in the UI, the CLI report and info.note.

Breaking

• Login is required by default. Set AUTH_USERNAME (a username or email) and AUTH_PASSWORD, or turn login off with AUTH_ENABLED=false. Without a password, one is generated and kept across restarts. Scripts sign in with HTTP Basic or a bearer token (AUTH_TOKEN).

Added

• Built-in login: signed session cookies, HTTP Basic, bearer token, and a rate limit on failed attempts.
• Video:
  • trim to a start and end time;
  • WebM (VP9 or AV1 + Opus) and AV1 in MP4;
  • video to animated GIF;
  • two-pass encoding for a target size in MB;
  • GPU encoders detected at startup: NVIDIA NVENC, Intel Quick Sync, VA-API and AMD AMF, next to Apple VideoToolbox.
• PDF compression with Ghostscript, with presets from "smallest" to "prepress". Ghostscript is included in the default image; -nopdf images leave it out.
• Compress in the browser. An optional mode using WebCodecs (Mediabunny), where files never leave the device. Anything the browser can't do falls back to the server, with a note.
• Download all results as one ZIP.
• Integrations:
  • live progress over Server-Sent Events;
  • webhooks signed with HMAC, with retries and SSRF protection;
  • batch status for many jobs in one call.
• Scale out: an optional Redis/BullMQ queue (QUEUE=redis), ROLE=web|worker, and a compress-media worker command.
• With object storage, inputs stay in the bucket, so workers on other hosts can fetch them.
• JSDoc types checked with tsc (npm run typecheck).
• THIRD_PARTY_NOTICES.md, and OCI license labels on the images (FFmpeg GPL, Ghostscript AGPL, libvips LGPL, Mediabunny MPL).

Fixed

• AVIF input was detected as HEIF.

Added

• Chunked, resumable uploads, modelled on S3 multipart uploads, for files of several GB. The part size (UPLOAD_PART_MB) and parallelism (UPLOAD_CONCURRENCY) are configurable, and interrupted uploads resume after a reload.
• Object storage: STORAGE=s3 works with any S3-compatible bucket (AWS S3, Cloudflare R2, Backblaze B2, DigitalOcean Spaces, SeaweedFS…). Browsers upload straight to the bucket with presigned URLs. There is a CORS check at startup, and S3_SETUP_CORS can set it up.

First public release.

• Web UI in English and Vietnamese: drag and drop, presets, before/after preview, redo with new settings, cancel, and download all.
• Video (MOV, MP4, MKV…) to H.264/H.265 MP4; images (JPG, PNG, HEIC, WebP, AVIF, GIF) to JPEG, WebP, AVIF or PNG; audio to MP3, M4A or Opus.
• CLI: compress, probe, info and serve, with a JSON report and folder mirroring.
• HTTP API for scripts and integrations.
• Docker image for linux/amd64 and linux/arm64, and Docker Compose files.
• Docs for people (README, docs/) and for AI agents (skills/, AGENTS.md, llms.txt).
• Tests: node:test for the API and CLI, and Playwright end-to-end tests on Chromium, Firefox and WebKit.