Know your browser’s boundaries.

Image tools

Use a current Chrome, Edge, Firefox or Safari with createImageBitmap and OffscreenCanvas in workers. JPG, still PNG and still WebP are accepted after checking their contents. The selected encoder must actually return the requested MIME type; unsupported WebP output is rejected instead of mislabeled.

HEIC, RAW, TIFF, SVG, GIF and animated PNG/WebP are outside this version. Images are decoded to ordinary canvas pixels: wide-gamut or HDR color and embedded ICC profiles can change. Image processing re-encodes content and does not preserve source metadata.

Memory safeguards

Ultra compression adds AVIF output and higher-effort JPG/WebP encoders, capped at 12 megapixels. It loads a single-thread WebAssembly encoder only when you process: about 3.5 MB for AVIF or under 0.3 MB for JPG/WebP, plus a small wrapper. No cross-origin isolation is required. AVIF can retain transparency, but Luma currently accepts only JPG, PNG and WebP as inputs. Check AVIF support in your destination app; Fast mode remains available.

AI Ultra uses Real-ESRGAN x4plus: about 67 MB of weights. It processes overlapping tiles at native 4× and resizes each tile to the requested output scale, avoiding a full-size 4× intermediate canvas. It uses the same local WebGL engine and is intended primarily for desktop testing. A heavier model can still smooth away wanted details or invent features.

The image upscaler offers 2× and 4× enlargement. AI mode needs WebGL graphics acceleration in a worker. All upscaling and enhancement outputs must fit 3840 × 2160, or portrait. Enhance at current size runs an AI pass blended gently with the original. Extra enhancement adds a separate pass after enlargement. Standard enlargement uses browser resampling. AI estimates detail; inspect the result, especially text and faces. AI Detail uses a larger model (about 29 MB per scale) for finer texture and can take several minutes. AI Fast uses a smaller model (under 1 MB per scale) with smoother results. Both also load an engine. Small phones may be slow or unable to run the models; try AI Fast or Standard if Detail is too demanding. The 30-minute processing timeout protects against stuck jobs.

These are conservative application boundaries, not a guarantee that every device has enough memory. Small phones may need much smaller files. There is no daily quota. ZIP creation, PDFs and video output can temporarily require several copies of the data.

Video beta

The compressor uses a local software engine (FFmpeg/WASM with x264) and exports H.264/AAC MP4. It loads only when compression starts, needs no cross-origin isolation, and can take several minutes. Browser playback support is separate from software decoding support. Resizing and conversion use the same local engine, with H.264 output in MP4, MOV or MKV. Audio extraction offers MP3, M4A, WAV or FLAC. Inputs are MP4, MOV, MKV and WebM. Playback support depends on your browser and device.

Every job checks codec support and rejects unexpected track loss. Only primary video and audio tracks are retained. The trimmer copies encoded packets and may retain footage outside the selected interval at keyframe boundaries. Resizing re-encodes; conversion can copy video already using the selected codec. These tools do not provide arbitrary format support, subtitle preservation or HDR guarantees.

When something goes wrong

Try a smaller file, a shorter clip or fewer images. Close other memory-heavy tabs. For video, try the other output format in a current desktop browser. If a browser stops a background tab, keep it in the foreground. Cancel releases the worker and allows a fresh attempt.

Desktop Chromium is the initial automated test target. Firefox, Safari and real iOS/Android devices require manual verification before a broad public compatibility claim. Report a reproducible issue.