HEIC photos that refuse to open on a colleague's laptop. A WebP download nothing wants to read. An EPUB you need on the printer by morning. This almanac collects the file conversions we field most often, the trade-offs each one hides, and the safest defaults when in doubt.
The image-format world fragmented twice in the last decade — once when Apple shipped HEIC by default in iOS 11, again when WebP became the de-facto format on the open web. Both shifts made files smaller; both made them harder to share. The four entries below close those gaps.
A decade after MIME types were supposed to make file extensions a polite suggestion, our inboxes still bounce IMG_4231.HEIC because Outlook can't show a preview. The right move is rarely "tell the sender to switch their phone settings." It's a clean conversion — and that's where most of the friction in this section originates.
HEIC encodes an image with HEVC, which is roughly half the size of an equivalent JPG at matching quality. The catch is licensing: Windows still ships HEIF support behind a paid extension, and most webmail previews refuse it outright. Converting to JPG at quality 92–95 produces a visually identical photo that opens on every device built since 1994.
WebP shaves 25–35% off comparable PNGs and is now served by roughly 96% of CDN edge nodes (CanIUse, Q1 2026). PowerPoint, older InDesign builds, and several CMS upload pipelines still don't accept it. PNG is the universal fallback: lossless, alpha-channel intact, and readable by every image library written since 1997.
The ICO container is older than the Web itself and has one trick worth remembering: it can hold a small atlas of bitmaps at different sizes, and Windows picks the best fit at draw time. A practical icon bundles 16, 32, 48, 64, 128 and 256 px — anything smaller looks blurry on the taskbar, anything larger is ignored by most renderers.
SVG is a markup file, not a picture, and that distinction matters once you paste a logo into Keynote and the gradients vanish. Rasterising at a known DPI gives you a fixed-size PNG with predictable rendering. Pick 2× the intended display size on retina displays; 300 DPI for anything destined for print.
PDF turned 33 this year. It remains the only document format you can hand to a stranger with confidence that the fonts, margins and pagination will not change. The four entries here orbit it: feeding documents in, taking pages back out.
An EPUB is, internally, a ZIP of HTML files. That makes it elegant on a Kindle and awkward on an office printer that wants page numbers and a margin. A good conversion freezes the reflowable text at a chosen page size (commonly A4 or 6×9 in), keeps the chapter table-of-contents as PDF bookmarks, and embeds the cover.
DOCX is OOXML — an Open Packaging Convention zip that references fonts by name, not by file. The most common reason an exported PDF looks "different" from the Word document is that the recipient's machine substituted a font for one of yours. A reliable conversion either embeds the typeface or swaps in a metrics-compatible match so pagination doesn't shift.
The most common reason to export a PDF page to JPG isn't archival — it's pasting page 4 of a contract into a Slack message without revealing pages 1–3. Choose 150 DPI for on-screen sharing, 300 DPI if it might be printed, and pick "selected pages" rather than "all" unless you really need an album.
Most accounting platforms accept PDF bundles but not photo dumps. Convert a sequence of phone snapshots into a single, ordered PDF: A4 or US Letter paper, margins of about 12 mm so the binder hole doesn't eat the timestamp, and don't recompress — JPG inside PDF is already as small as it's going to get.
Three lossless conversions for the people who spend their days moving structured data between tools that disagree about how to spell "true". None of them require uploading anything anywhere — they run entirely on your machine.
The headers become keys, each row becomes an object, and every well-behaved CSV-to-JSON converter has to make three quiet decisions: how to detect the delimiter (semicolons appear in 14 of the 27 EU locales), how to type values (the string "0123" is usually a zip code, not a number), and whether to coerce "" to null. Pick the defaults that are safe for your downstream tool, not the most "clever".
The reverse trip is harder, because JSON has nested objects and arrays that CSV does not. The dotted-key convention — address.city, items[0].sku — is what BI tools expect and what Excel can still parse without macros. Decide up front whether arrays of objects fan out into multiple rows or collapse into a JSON-encoded cell.
XML survives in SOAP responses, RSS feeds, sitemaps, and a stubborn segment of banking and insurance APIs. The conversion is mostly mechanical, with two exceptions: XML has both attributes and child nodes (JSON does not), and XML allows mixed content where text and elements interleave. The pragmatic default — attributes prefixed with @, text under #text — is ugly but lossless.
A common misunderstanding: an MKV-to-MP4 conversion is rarely a video conversion. The video inside both files is usually identical H.264 or H.265. What changes is the wrapper.
Matroska (MKV) is a more permissive container — it accepts more codecs and more subtitle formats than MP4. But Safari, QuickTime, most TVs and most chat clients prefer MP4. When the inner streams are already H.264 or H.265 video with AAC audio, the right operation is a remux: the codec data is copied byte-for-byte, only the container is rewritten. A two-hour film completes in seconds and loses no quality.
WAV is PCM — uncompressed audio at roughly 10 MB per minute. MP3 at 192 kbps is roughly 1.4 MB per minute, indistinguishable from the source in blind tests for non-classical material. Use 320 kbps if you'll re-edit later, 192 kbps for distribution, 96–128 kbps for spoken word. Preserve ID3 tags — every podcast platform needs them.
A single table to bookmark. "Local" means the conversion can run entirely on your machine; "Server" means a brief round-trip with a one-hour retention policy.
| Cat. | From | To | Where it runs | Batch | Lossless |
|---|---|---|---|---|---|
| Img | HEIC | JPG | Local | Yes | ~ visual |
| Img | WebP | PNG | Local | Yes | Yes |
| Img | PNG | ICO | Local | Yes | Yes |
| Img | SVG | PNG | Local | Yes | Rasterised |
| Doc | EPUB | Server (≤1h) | Yes | Reflow → fixed | |
| Doc | DOCX | Server (≤1h) | Yes | Yes | |
| Doc | JPG | Local | Yes | Rasterised | |
| Doc | JPG | Local | Yes | Yes | |
| Dat | CSV | JSON | Local | Yes | Yes |
| Dat | JSON | CSV | Local | Yes | Flattened |
| Dat | XML | JSON | Local | Yes | Yes |
| Vid | MKV | MP4 | Local (ffmpeg.wasm) | Yes | Stream-copy |
| Aud | WAV | MP3 | Local (ffmpeg.wasm) | Yes | Lossy |
A small selection of the most frequent reader letters, with answers checked against the latest specifications. Have a question we haven't covered? Send it in — we revise this section each quarter.
No, but the loss is negligible at quality 92 and above. HEIC itself was decoded losslessly from HEVC; the loss happens when the pixels are re-encoded into JPG's DCT scheme. For photographs, the difference is below the threshold of perception on calibrated displays [1]. If you need pixel-perfect output for archival, convert HEIC to PNG instead.
That's expected. WebP is a compressed format with a built-in psychovisual model; PNG is mathematically lossless and stores every pixel. A 250 KB WebP typically expands to a 700–900 KB PNG. You're trading file size for compatibility — that's the whole point of the conversion.
Yes. The bundled sizes (16 / 32 / 48 / 64 / 128 / 256 px) match Microsoft's published guidance for executable resources and Start menu tiles [2]. Visual Studio's resource compiler will pick them up directly.
You set the output width. The SVG renderer rasterises at exactly that size, so the result looks crisp at the chosen resolution and slightly soft when scaled up further. For retina-class displays, export at 2× the intended display size; for print, 300 DPI is the safe default.
Yes — numbers, booleans, ISO-8601 dates, and null-like sentinels (NA, N/A, empty string) are coerced. You can disable detection to keep everything as strings, which is the safest mode for downstream parsers that handle their own coercion.
Yes, using dotted keys (address.city). Arrays of primitives are joined with a separator you choose (default: |). Arrays of objects fan out into multiple rows by default, but you can switch to JSON-encoded cells if you want one row per input record.
Embedded fonts are passed straight through to the PDF. For non-embedded fonts, the converter substitutes a metrics-compatible match — same character widths, so line breaks and pagination don't shift. This is the same strategy Microsoft Word uses when "embed fonts" is enabled at save time [3].
Only when it must. If the MKV's video stream is H.264 or H.265 and the audio is AAC, the operation is a remux: the streams are copied byte-for-byte into an MP4 container [4]. The result is bit-identical to the source. Re-encoding kicks in only for legacy codecs (VP9 audio, FLAC audio, Theora, etc.).
320 kbps for music you intend to re-edit; 192 kbps for general listening (the long-standing public-broadcast default); 128 kbps for spoken-word podcasts. Our recommendation, in line with the AES guidance on perceptual audio coding [5], is to default to 192 kbps and step up only when re-encoding is anticipated.
Most of the conversions in this almanac never upload anything — image, data, audio, and video work happens inside your browser using WebAssembly. EPUB-to-PDF and DOCX-to-PDF do round-trip to a server because of font handling; those files are deleted within sixty minutes, never indexed, and never used to train any model.