Changelog

Version 2.2.1.0

Fixed

• The AI features now work on a clean Windows install. If you installed ImageLabs on a computer that has never had developer tools on it, then background removal, colorize, repair, AI 4× (both engines), blur by distance and face detection would all fail to start — while the app itself launched normally and every ordinary edit (resize, convert, crop, icons) worked exactly as it should. The packages for 2.0, 2.1 and 2.2 left out four small system libraries the AI engine needs; they happen to already be present on a developer’s machine, which is why the builds tested fine and shipped anyway. They are now included. Nothing to install and nothing to configure: the same download simply works.
• If you hit this, what you most likely saw was a model failing to load right after a download finished — which reads like a broken download but was not one. The download was fine, and re-downloading would never have helped.

Version 2.2.0.0

Added

• Hide faces in Blur / Mosaic. Area gains Faces, which finds everyone in the photo and mosaics or blurs each of them — and Keep one face, where you click the person to leave alone and everyone else is hidden. That is the usual real need: publishing a picture of your own child at an event without exposing the other children in it. The panel says how many faces it found, including none, because a redaction that quietly did nothing is the one failure you must not discover after publishing. Like the face crop it is recorded as a step that re-checks each photo, so it works across a folder.
• Crop to faces. In Crop, frame the photo on the people in it — everyone, or just the main face — as a square, 4:5, 16:9, or however it lands, with the eyes placed on the upper third the way a portrait usually wants them. Unlike an ordinary crop this is recorded as a step that measures each photo afresh, so replaying it over a folder gives you one properly framed portrait per picture instead of the same rectangle cut out of all of them. A photo with nobody in it is left untouched rather than cropped to a guess.
• Watermarks can place themselves away from faces. Position gains Auto — avoid faces, which finds the faces in the photo and puts the mark in whichever corner they leave clear. The existing Auto — avoid subject covers the whole person, so on a portrait it finds every corner equally bad; a face is small and unambiguous, and the mark is free to sit over a shoulder or the background. It keeps the corner you picked unless a face is actually in the way, and tells you how many faces it found — including when it found none, so “nothing moved” never has to be guessed at. Face detection is included in the app; nothing to download.
• Blur by distance instead of around a cut-out. Background → Blur gains a Falloff choice: the existing Behind the subject (cut-out), or By distance (depth), which works out how far away everything in the photo is and lets the blur grow with distance. A cut-out blur puts one single blur everywhere behind a hard edge, and that edge is what gives it away — hair, railings, a hand held out, anything the mask gets slightly wrong sits against a wall of even blur. A distance blur has no edge to get wrong: things a little further back go a little softer, so the photo keeps its depth instead of separating into a subject and a backdrop. Keep sharp chooses which distance stays in focus — left at nearest it is the ordinary portrait; moved further in it focuses past the foreground, so whatever is in front of your subject goes soft too, the way it does through a real lens. It needs no subject at all, so nothing is detected and the cut-out models are never downloaded for it.
• A second AI enlarger, for drawings. The engine row in Resize is now Standard, AI 4× — photo, and AI 4× — anime / line art. The photo engine is trained on photographs, and on a drawing it does what it was taught to do: it treats the flat colour as texture to invent detail in, and softens the linework that carries the picture. The art engine keeps the lines hard and the fills clean. It is also the smaller download of the two — 18 MB against 67 — and noticeably quicker. Like the photo engine it downloads on first use and shows you the size before it does anything, and each engine has its own download, so choosing one does not fetch the other.
• Batch says which photos have no face in them, before it writes anything. A step that hides or crops to faces does nothing at all to a photo where none was found — it gets written out looking processed, with everyone in it still recognisable. Across a folder nobody checks each file, so the discovery happens after publishing. When the recorded steps need faces, the folder check now looks at every image and reports what it found: “2 of the 5 images have no face in them — they would be saved with nothing hidden”, or that a face was found in all of them. It warns rather than blocks, and only runs when the chain actually needs faces.
• The status line says which step is running. A long edit used to report only “Running AI…” from beginning to end, which answers neither of the questions you actually have while waiting: what is it doing, and why is this one taking longer than the last. It now names each slow step as it starts — finding the subject, estimating distance, blurring by distance, colorizing, repairing, upscaling. Steps answered from a cache stay silent, so the first edit of a photo says it is estimating distance and nudging a slider afterwards does not — which is also the reason the second one was quicker. There is no percentage, because there is no honest one to show: the models report no progress at all, so any bar would be a guess dressed up as a measurement. What you get instead is a small moving bar beside the text, and only after half a second of waiting.

Fixed

• Blurring the background is no longer a coffee break. Choosing Background → Blur on an ordinary photo could sit there for over a minute, on every preview and again on Apply, with nothing to show that it was working rather than stuck. The blur was being computed in a form whose cost climbs with the square of the radius: 69 seconds on a 2048 px photo at the default strength, and minutes at the top of the slider, where the wait got worse the stronger you asked for. It now lands in about a second. The picture is unchanged — a blur is exactly the operation that cannot tell the difference, and the sharp subject is composited back over it either way. The Blur / Mosaic tool’s own blur slider was on the same path.
• The Batch panel keeps up with Undo and Reset. The recorded edit chain, the Apply and Save-recipe buttons and the folder summary only refreshed while Batch was the task on screen, so undoing or resetting from another task left a chain that no longer existed still listed and its buttons still live. They now refresh with every change to the chain, and the folder summary re-checks which images differ in size when undo changes the size of the photo you edited.
• Task names in the rail are no longer cut off. “Repair / Scratch” and “Compress / Convert” were sliced mid-word with no ellipsis to show it. The rail is wide enough for every name, and a name that ever outgrows it will show an ellipsis rather than being silently clipped.
• The batch folder summary reads correctly for a single image — “1 is a different size” rather than “1 are a different size”.
• Recipe files no longer carry a derived copy of the selection points, which could only ever be ignored on load.

Version 2.1.0.0

Fixed

• Repair / Scratch removal only touches what you painted. It was replacing the whole photo with a low-resolution version of itself instead of filling just the painted damage — the brush mask reached the engine carrying an opaque transparency channel, which the compositor preferred over the painted strokes, so the strokes were ignored. Unmasked pixels changed: 100% → 0%.
• Colorize keeps transparency. Removing a background and then colorizing the cut-out filled the transparent area back in solid; the transparency channel is now carried across the colorization step.
• AI models are released from memory. Trying several models in one session grew the app to about 4.4 GB and nothing was ever given back — on smaller machines that meant heavy paging or an out-of-memory failure in an unrelated task. Models are now held in a size-bounded cache that releases the least-recently-used ones, and a model in use is never unloaded mid-operation. Peak use 4.4 GB → 2.6 GB, and AI Models gained a Free memory button that drops back to about 100 MB.
• Batch adapts to each image's size. Crop, Blur/Mosaic and Add Text were replayed at fixed pixel positions, so any image not the same size as the one you edited was cropped in the wrong place at the wrong scale — silently, across the whole folder. A chain recorded on a 3000×2000 photo turned a 640×480 file into 140×80; it now produces the correct 427×288. Text and watermark sizes scale to match, and a crop can no longer fall outside a smaller image and produce a sliver.
• Object picking is much faster after the first click. Clicking a second object in the same image re-ran the heavy analysis pass every time; it is now computed once per image and reused — 1282 ms → 69 ms per subsequent click.
• Colorize and Repair report a missing model instead of starting a silent, uncancellable multi-hundred-megabyte download with no progress — reachable by replaying those steps in Batch after deleting the model.
• Saving to a path with no folder component no longer fails.

Added

• Compress can hit a size you name. Type a target under Fit to a size in Compress / Convert and ImageLabs finds the highest JPEG or WebP quality that stays under it, instead of leaving you to hunt with the slider. If the target can't be reached at the image's current resolution it says so, rather than saving something too big.
• Refine an object selection with more clicks. In Background and Blur/Mosaic, a plain click still picks an object exactly as before; Shift-click now adds another part to the selection and right-click leaves a part out. Extra clicks are nearly instant thanks to the caching above, and the refined selection is saved with the edit so undo and Batch stay exact.
• Save and reuse an edit recipe. The edits you've applied can be written to a small file and applied to another photo later, or shared. Loaded steps appear in History and undo normally, and their crops and positions scale to the new photo's size. Repair steps are left out — a painted mask only means something on the photo it was drawn over.
• Batch tells you what it's about to do. Choosing a folder now reports how many images were found, how many differ in size from the one you edited, and whether a previous run's output is about to be replaced. Preview first result applies your edits to a single image so you can look at the outcome before running the folder.

Changed

• Repairs are made at full resolution. Repair used to shrink the entire photo to the model's fixed 512-pixel input, so on a large image the filled area came back as a soft, upscaled patch that never quite matched its surroundings. Painted areas are now grouped and each is repaired at its own scale, so a scratch in a 12-megapixel photo is reconstructed at full detail — and separate scratches are handled separately, so fixing four marks across a photo costs no more quality than fixing one.
• Imaging engine updated, clearing 29 known security advisories in the component that reads image files. Verified to produce byte-identical output across every operation and output format.
• AI Models now shows how much memory loaded models are using, next to how much disk the downloads take.

Notes

• Existing behaviour is unchanged: a 33-case regression suite covering every editing operation and all seven output formats produces byte-identical results before and after every change in this release, including the engine update.

Version 2.0.0.0

Added

• Nine on-device AI models — a managed, offline library. Every model runs locally with GPU acceleration and automatic CPU fallback, is integrity-verified (SHA-256), and is open-source. One is bundled; the rest download on demand:
  • Background removal (5) — fast (bundled, 4.7 MB) · balanced (43 MB) · best (176 MB) · people (176 MB) · sharpest edges (176 MB), all Apache-2.0.
  • Click-to-select any object (43 MB, MIT).
  • AI 4× super-resolution (67 MB, BSD-3-Clause).
  • Colorization of black-and-white / faded photos (935 MB, Apache-2.0).
  • Paint-to-fill repair / inpainting (88 MB, Apache-2.0).
• Background task with four AI modes (was “Remove Background”) — Remove, Blur (portrait/bokeh), Color pop (color subject on a grayscale background), and Replace with image. Remove mode gains Trim to subject, Sticker outline (width + color), and Drop shadow.
• Click-to-select object — instead of the auto-detected subject, click the exact object you want. Powers Background → Pick object, Crop → Select object, and object-shaped Blur / Mosaic.
• AI 4× upscaling in Resize — choose the AI enlargement engine when making an image bigger; it runs tiled (seam-free), fully offline, keeps transparency, then fits your exact target size.
• Colorize task — add natural, realistic color to black-and-white or faded photos with one click; original detail is preserved and only color is generated. Works on true grayscale files, not just color photos that look gray.
• Repair / Scratch removal task — paint over scratches, tears or small unwanted objects with an adjustable brush, then fill them in; only the painted areas change, with undo-per-stroke and clear.
• Extract Text (OCR) task — recognize text in an image fully offline, with a language picker (follows installed Windows language packs), an editable result box, Copy, and Save .txt.
• Replace Color task — the image's dominant colors are detected as clickable swatches; pick a color from a swatch, a hex value, or the eyedropper, set a tolerance, and preview live — including a Transparent plate that erases the picked color.
• Blur / Mosaic task — drag a selection (or click an object) over faces, plates or addresses and pixelate or blur them, with live preview.
• Add Text task — click anywhere on the image to place text (size + color).
• Compress / Convert task — pick a format and quality with a live file-size estimate before saving.
• Metadata task — inspect EXIF and image facts; pairs with a Strip EXIF on save toggle.
• Batch task — replays this image's edit chain (everything in History, AI steps included) over every image in a chosen folder, with progress and cancel; results land in an imagelab-out subfolder.
• AI Models manager — see every model's status, license and size; download, delete, or import a model file on offline machines (integrity-checked); open the cache folder.
• Redesigned workspace — the task rail is grouped into EDIT / AI & COLOR / RESTORE / ANNOTATE / TOOLS, with a new Overview showing Quick actions, Image info, and a live History of applied edits.
• Auto enhance in Adjust (automatic levels + gamma), and a Watermark “Auto — avoid subject (AI)” position that places text where the subject covers least.
• AI mask caching — the model runs once per image state, so switching modes and dragging sliders re-composite instantly.
• Preview zoom & pan — zoom around the cursor with Ctrl+scroll, drag to pan, double-click to toggle Fit ⟷ 100%, with an on-preview −/+/Fit/1:1 toolbar.
• Redo and keyboard shortcuts (Ctrl+O/S/Z/Y), paste / copy image (Ctrl+V / Ctrl+C), hold-to-compare, a split Save button (export again / overwrite original), a modern color picker on every color input, unsaved-changes protection, and a single-instance guard.

Fixed

• Colorize now works on true black-and-white files — the generated color is no longer discarded when compositing back onto a grayscale image; the working image is promoted to full color first.
• EXIF orientation is now honored — photos from phones/cameras no longer load or save sideways.
• Crop overlay tracks window resizes — resizing the window in Crop mode no longer shifts the crop mapping.
• Preview/edit hardening — rapid slider changes, opening a new image mid-preview, or closing mid-render can no longer race the image engine (previews are serialized, latest-wins).
• ICO export no longer distorts non-square images — frames are fit and padded with transparency instead of squashed square.
• Large-image responsiveness — open, apply, rotate/flip, undo and reset run off the UI thread; the window no longer freezes on big photos.
• AI model downloads are integrity-checked — SHA-256 verified, size-validated and serialized; a corrupt or intercepted download can't be cached as a model.
• HEIC files can be picked in the Open dialog, animated/multi-page images warn on open, and undo history is bounded to the last 10 steps.
• Saves are atomic — closing mid-save no longer truncates the output; the window drains in-flight work before closing, and closing responds on the first click.
• Errors no longer close the app silently — unexpected failures show a message and are logged; model-status text follows the light/dark theme.

Version 1.0.0.0

Added

• Task-based editor — open one image and apply tasks from the left rail: Crop, Resize, Rotate / Flip, Remove Background, Adjust, Watermark.
• Chainable edits — each Apply commits to a working image so edits stack (e.g. crop → remove background → resize), with Undo and Reset.
• Live preview — the result updates as you change a task's settings; Apply keeps it.
• Interactive crop — drag the selection box and corner handles on the preview, with a live pixel-size readout.
• Resize — by width/height, keep-aspect, and don't-enlarge.
• Rotate / Flip — 90° left/right and horizontal/vertical flip (applied immediately).
• Remove Background — an offline AI smart cut-out with a fast bundled model and an optional best-quality download (with progress and a ready indicator), GPU acceleration with automatic CPU fallback, edge feathering, and a fill-background-with-colour option — plus a manual flat-colour removal with adjustable tolerance.
• Adjust — brightness, contrast, grayscale, sepia, blur, sharpen.
• Watermark — text with position, font size, and opacity.
• Save result — choose the format at save time: PNG, JPEG, WebP, BMP, TIFF, GIF, and multi-size ICO; quality control for lossy formats.
• Light / dark theme toggle (remembered between launches).
• Branded experience — app icon, startup splash, and an About box with a link to backendside.com.
• Self-contained packaging — a Microsoft Store-ready package with no prerequisites to install.

Notes

• Inference is 100% local — no account, no cloud, nothing leaves the PC.