Changelog

What we shipped, what we fixed

Dated receipts. No marketing fluff. Things break and get fixed openly.

Newest first. We update this when something meaningful changes, not on a schedule.

2026-08-28

We said "entirely offline" across the tool pages. That was true of the processing, not of loading the page. Eighteen PDF/document tool pages and several image tool pages had "Offline" or "entirely offline" in their title or hero copy, wording we ourselves added four days earlier for search visibility (see 2026-08-24, below). What's actually true, and what the pages say now: the work happens on your device, but PixThisApp is a website, and loading it the first time still needs a connection like any other site. We'd rather lose the phrase than keep a claim that isn't accurate.
Image to Text and PDF to Text now say what they actually download. Both tools' OCR engine and language data, about 7MB, come from a CDN the first time you use either tool on a given device. We measured the real number rather than guessing, and the pages now say so plainly. Your file itself is never uploaded; only the OCR engine comes from outside your device.
Batch compress-to-target-size doesn't tune each file individually, and we've fixed the copy that implied it did. It picks one quality level for the whole batch and applies it to every file, aiming for the group's total size, not a per-file optimum. Corrected on the tool page and inside the tool itself.
"Zero-Knowledge Architecture" was the wrong term on our security page, so it's gone. That names a specific cryptographic property we don't implement, and don't need to: nothing leaves your device at all, no proof-of-knowledge required. We also stopped calling our own internal review an "audit", it's a self-review, not an independent one, and the page now says so.
Corrected "no cookies, no analytics": we do run one, an anonymous page-count only. Our privacy page and both comparison pages said we use no analytics at all. We do, in the narrowest form available: Cloudflare's cookieless aggregate page-count, with no per-visitor tracking. Copy now matches what's actually running.
Corrected which engines actually run PixThisApp. Our tech page named "Rust" as one of the compiled engines; none of our WebAssembly modules are Rust. It's C and C++ (QPDF, Tesseract, ONNX Runtime) plus JavaScript (pdf-lib). It also implied GPU inference; WebGPU is off in production, everything runs on your CPU through WebAssembly.
Removed a benchmark comparison on Compress Image we can't back up. The page claimed output at "TinyPNG quality." We have never actually benchmarked against TinyPNG, so that specific comparison is gone, replaced with what we can stand behind: smaller files, no upload.
Fixed dark-page header text that was close to unreadable. On the roughly 33 tool, comparison, and changelog pages with a dark background, the top navigation links were still styled for a light page (near-black text), making them close to invisible. They now switch to a light, high-contrast style on dark pages, measured at 4.5:1 contrast or better against the real rendered background, at both desktop and mobile widths.

2026-08-24

PixVoice: returning visitors were getting a 300-700MB download with no warning. Every return visit to the voice workspace silently re-triggered the model download for your saved setup the moment the page loaded. First-time visitors already got a one-tap consent screen for this; returning visitors didn't. They now see a "welcome back" screen naming the saved setup and its real size before anything downloads.
PixVoice workspace didn't line up with its own header. The header, the working area, and the footer on the voice workspace page were three different widths, because an earlier site-wide layout fix only reached the shared components, and this page has its own separate header and footer. All three now share the same edge.
Share button added to the last 4 tools that were missing it. Image to Text, PDF to Text, PDF to CSV, and Photo to PDF were the only 4 of 24 tools without the one-tap share button the rest already have. Same pattern, same privacy note, nothing new invented.
Two real security holes closed. We updated the PDF rendering library (pdfjs-dist) to close a flaw that could let a crafted PDF run arbitrary code just by opening it, and pinned a nested dependency (protobufjs, pulled in through our image-AI engine) that carried a critical remote-code-execution flaw of its own. Neither was theoretical: both were reachable through normal use of the app.
Added the word "offline" back into 18 PDF/document tool page titles, for search visibility. This is the change reversed four days later (see 2026-08-28, above), once it became clear "offline" wasn't an accurate description of a page that still needs a connection to load.

2026-08-15

The opening animation has been rebuilt. The dot now arcs in on a single curve, lands on the letter it belongs to, bounces twice and settles, and the wordmark carries proper depth and shadow. It is drawn entirely by the browser rather than played as a video, which means it costs about six kilobytes in total and stays sharp on any screen size. Phones still skip it, as they have since May, because an interstitial is the fastest way to lose someone on a weak signal.
The logo in the corner now does something. It always had a click handler, but that handler reloaded the page while the animation was suppressed, so it looked broken to anyone who tried it. Clicking it now replays the opening and returns you to the homepage, from wherever you are. It works from the keyboard too.
New favicon, and the old light/dark version turned out to not really be switching. The previous icon was close to invisible in a dark browser tab, and the orange in our mark didn't survive its export at all. The replacement is clearly visible on light and dark tab bars alike, so the old "switches with your system theme" version is retired rather than duplicated. That switch was mostly not doing anything to begin with: outside the two smallest icon sizes, the light and dark files were identical, so you'd have gotten the same icon regardless of your system setting.
Homepage header now lines up exactly with the tools grid below it. Its left edge used to drift slightly from the grid's as the window got wider, because it used a different padding rule. Fixed, and checked across 10 screen widths; the gap is exactly zero at every one now.
PixSign, still in development, is no longer indexable. It went live inside a larger push on 12 Aug without us noticing it was crawlable. It now carries a noindex tag and its sitemap generation is disabled. The main site's own listing is unaffected.
Correcting our own number: the intro splash weighs about six kilobytes, not one and a half. An earlier version of this entry counted only the font subset. Measured properly, script plus inline pre-paint code plus font, compressed, it's about six kilobytes.

2026-08-12

Clicking during our opening animation could open a tool you never picked. The orange screen that plays when the homepage first loads covered everything, but it did not actually catch clicks. Anything you clicked while it was up went straight through to the tool grid underneath, so you could find yourself inside Remove Background, or any other tool, without having chosen it. It behaved that way since 2 May, the day we built that screen, until today. The opening screen now blocks clicks while it is on screen and releases them the moment it finishes.
We also wrote a test for the worse version of this bug. A screen that blocks clicks is only safe if it reliably stops blocking them, and the failure nobody notices until it is live is the opposite one: an opening animation that never lets go, leaving the whole site unclickable. There is now a committed test that fails if the screen stops blocking, and a second that fails if it stops releasing. Both are checked against a deliberately broken version first, so we know they can actually fail rather than passing out of habit.

2026-08-11

Compress PDF was leaving real savings on the table. Plenty of scanners record a page's size wrongly, writing pixel dimensions into the field meant for inches. We trusted that number, so on those files we were drawing each page around three times larger than the scan actually inside it, then compressing that enlargement. Files still got smaller, but the file we handed back could be roughly twice the size it needed to be, and the tool appeared to hit a floor it could not pass. The fix: we never draw a page larger than the images actually on it, because enlarging a scan cannot recover detail that was never captured. The same documents now compress properly and finish about four times faster, since we stopped spending that time drawing pixels we invented.
And in one case it did nothing at all, while saying it had succeeded. If you fed it a file whose pages were already fairly small, we would draw them larger than they really were, the result would come out bigger than what you started with, and a safety check would quietly hand your original back labelled "already optimal". It was not optimal. It was untouched. We have measured one real document where the file came back byte-for-byte identical with that badge showing. If that happened to you, the same fix resolves it, and we are sorry it said otherwise.
The same flaw was in three other tools, and is fixed in all of them. PDF to JPG, Sign PDF and our text extraction shared the code that decides how large to draw a page. On documents with a mis-recorded page size, all of them were building enormous images: at 300 DPI, PDF to JPG was drawing a single page at over 150 megapixels. Browsers have limits there, and past them things fail quietly rather than loudly. All four tools now go through one shared guard.

2026-07-20

We were claiming GPU acceleration we do not do. The FAQ said Chrome and Edge run the AI tools faster because of WebGPU, and that Safari and Firefox fall back to something slower. None of that is true. Every tool runs the same WebAssembly code on your CPU in every browser. There is no fast browser, no slow browser, and no fallback tier. We also said processing uses "your CPU and GPU"; it is CPU only. We did try the GPU path, and it crashed on real machines, so we chose the slower option that always works.
Upscale Image is not AI, and we stopped implying it was. It was listed as one of our AI tools. It is not: it uses high-quality bicubic resampling. We shipped a neural upscaler, but it ran out of memory on real devices, so we removed it in May and kept the method that finishes. The tool page and FAQ now say this outright.
PixVoice: karaoke captions are word-exact on desktop, approximate on phones. The caption page promised Whisper word-level alignment without mentioning that phones do not get it. They do not: the aligner is a 400MB model that exhausts memory on most phones, so mobile falls back to timing estimated from the speech itself. Close, but not word-exact. Since that page is aimed at people making TikToks, and most of them are on a phone, saying nothing was the wrong call. There is now a question about it near the top rather than a footnote at the bottom.
PixVoice video export: MP4 on Chrome and Edge, WebM on Firefox. We said MP4, full stop. Firefox cannot record MP4, so it gets WebM, which some platforms want converted before upload. Said plainly now.
PixVoice: hear the voices before you download anything. The landing page now plays real samples of five of the 28 voices, each saying something different so you can tell them apart. They are genuine Kokoro recordings made with the same engine that runs in your browser, pre-rendered so that listening costs you nothing: no model download, no server, and the audio only loads when you press play. We also corrected the download size on that page. It said 700MB flat; text to speech alone is about 300MB, and 700MB is only if you add karaoke captions.
The same "offline" correction, now across every tool page. Fifteen tool pages still described themselves as working offline, and the FAQ still promised full offline support was coming. Neither was true, so both are gone. What those pages say now is what actually happens: the work runs locally in your browser and your files are never uploaded. Loading the site still needs a connection, and we say so rather than implying otherwise.
We removed an "offline" claim we hadn't earned. The download page and the homepage said PixThisApp works offline after you install it. It doesn't yet: there is no service worker, so loading the site still needs a connection. What IS true, and what we now say instead: your files are never uploaded and the processing runs on your device. The site and the code each tool needs still load from the web, and AI tools download their model the first time you use them. Installing still gets you a real app icon and its own window. We'd rather correct ourselves than let a promise sit there unearned.

2026-07-13

Unlock PDF is back, for real this time. It now removes a PDF's password using a genuine decrypt engine (QPDF, compiled to WebAssembly). You enter the password you know, it unlocks on your device, nothing is uploaded. This replaces the paused version, which relied on a library that couldn't actually decrypt.
Protect PDF shipped. Add an AES-256 password to any PDF so it needs that password to open. Same QPDF engine, same rule: the file and password never leave your browser. This clears the last "coming back" PDF promise.
Unlock PDF tells the truth when a file isn't locked. Drop a PDF that has no password and it now says so plainly instead of claiming it "unlocked" something. Owner-restricted PDFs (open fine but block printing/copying) still unlock with an empty password.
Share your result in one tap. After a tool finishes, the Share button now hands off your file to WhatsApp, email, or anywhere your device shares, with a short note that it was made free and private on pixthisapp.com and never uploaded. We never watermark or alter your file, and the note is text you can delete before sending.

2026-07-10

Site-wide honesty pass. Went through the tool copy and removed claims we couldn't stand behind — quality-loss wording on compression, "AI upscale" where it isn't AI, bookmark/hyperlink preservation the PDF library doesn't actually keep, a phantom PDF-to-Word on the comparison page, and WebGPU/voice-count claims on PixVoice. If it's on the page, the tool does it.
Privacy Blur — real redaction. Black-out mode now permanently destroys the covered pixels (irreversible, verified), with honest labels: blur/pixelate are marked as reversible style effects, not privacy.
Correctness pass across the tools. Filenames and format badges now match the file you actually get; Compress Image + Compress PDF keep your original when "compressing" would make it bigger (no fake savings); PDF to Text spaces words correctly; PDF to CSV column detection improved with an honest "verify amounts" note; Crop straighten and Sign PDF now line up on rotated pages; Strip Metadata verifies its own output. Still no uploads.
Unlock PDF paused. Our PDF library can't actually decrypt a protected PDF, so we removed the tool rather than fake it. It's on the "coming back when fixed" list and returns with a real decrypt path (QPDF-WASM).

2026-05-10

Remove Background — new AI engine (Desktop only). Replaced RMBG-1.4 with BiRefNet_lite on capable desktop devices. Same architecture family as the state-of-the-art RMBG-2.0, but publicly accessible. Edge accuracy jumps from ~74% to 90%+. Inference is noticeably faster — most images complete in 5–10 seconds. Mobile devices will continue to use the lighter legacy engine for stability. Also fixes the intermittent "washed out mask" issue caused by an upstream ONNX Runtime bug we were silently hitting.
PixVoice — Neon Pulse caption style redesigned. Active word now has a bright white core with intense neon glow and a subtle scale-bounce animation. Past words are solid orange, future words are crisp white. Much more legible on busy backgrounds.
PixVoice — premium background presets. Flat 2-color gradients replaced with a new volumetric mesh gradient system — layered radial light orbs over dark bases. All 6 presets (Carbon, Obsidian, Hermès, Ocean, Sunset, Aurora) now look studio-grade.
PixVoice — video background uploads. You can now upload MP4/WebM video files as custom backgrounds for video export. Live-looping preview swatch in the UI.
PixVoice — generation stability. Fixed a race condition where starting a new session while the WASM engine was still running could deadlock the UI. Controls are now disabled during active generation.
Mobile download compliance. Replaced a fragile programmatic download script that was causing mobile browsers (iOS/Android Safari & Chrome) to silently block downloads across all 23 tools. All tools now require a manual, explicit click on the orange "Download" button to save files, strictly adhering to browser security models.

2026-05-03

Comparison pages live — `/compare/pixthisapp-vs-ilovepdf`, `/compare/pixthisapp-vs-iloveimg`, `/compare/pixthisapp-vs-camscanner`. Honest side-by-side, including where the competitor wins.
Hero copy refreshed on the homepage — leans on what's verifiable in DevTools, not adjectives.
CLS fix — homepage hero text no longer pushes the tools grid down during load.
Tier-aware performance notice on landing — mid/low-end mobile users get an honest heads-up that AI tools may be slow.
"New" badges removed from 8 tools (Photo to PDF, Sign PDF, Delete Pages, Compress PDF, PDF to Text, Image to Text, Unlock PDF, QR Generator). They weren't new anymore. Going forward: "New" only on tools shipped in the last ~30 days.

2026-05-02

Image build corruption resolved. Some AI image tools were failing for a stretch of about two weeks. Workers + WASM bundles rebuilt; AI tools (Remove Background, Upscale) verified working end-to-end. If you visited during the broken window, sorry — it's fixed.

2026-04-13

Tools parked to legacy: Bento Collage, Asset Engine, Rename, Color Palette. Low usage; revival pending if real demand returns. We don't quietly delete tools — they live in legacy with a path back.

Earlier

Coming back when fixed: Object Remover (LaMa model upstream removal, pending replacement) and Remove Watermark (DeepFillV2 ONNX WASM error). We don't pretend they exist while they don't. (Unlock PDF and Protect PDF are no longer on this list - both shipped 2026-07-13 on the QPDF engine.)
Beta capability: PDF to CSV is labeled "Beta" because table extraction quality varies by document. Honest disclosure on the tool itself.

How we work

PixThisApp is built and improved by a small team. Updates ship when they're ready, not on a marketing calendar. Sometimes things break — when they do, we fix them openly and document it here.

If a tool you rely on stops working, refresh the next day or check this page. Most issues are resolved within 48 hours.

We document changes here because "trust us, we improve" is a claim. A dated changelog is evidence.

Open PixThisApp