The transparency-fixer now directly detects the baked-in checkerboard
pattern using per-pixel chroma analysis instead of BiRefNet AI matting.
Achromatic pixels in the gray range are classified as background
(transparent), chromatic pixels as foreground (opaque), with smooth
transitions at anti-aliased edges.
- No longer requires Python sidecar or background-removal bundle
- Watermark removal uses Sharp median(5) filter pre-processing
- Moved tool from "ai" to "utilities" category
- Removed from PYTHON_SIDECAR_TOOLS and background-removal enablesTools
- Near-instant processing (pure Sharp, no model inference)
The luminance anomaly detection + LaMa inpainting approach failed because
watermark signal on the matted foreground was too weak (10-15 units vs
threshold of 25). Median filter with kernel=5 effectively removes
semi-transparent watermark text while preserving the stamp structure.
Pipeline is now: median filter (if toggle on) -> BiRefNet matting -> defringe.
No longer requires object-eraser-colorize bundle for watermark removal.
Evidence-based spec informed by diagnostic testing on 4 sample images.
Addresses catastrophic scratch over-detection (up to 68.7% false
coverage on small images), LaMa 512x512 resolution loss, CodeFormer
over-smoothing on small faces, and excessive NLMeans defaults.
Add tier enum (fast/balanced/high) with balanced default to the Zod
settings schema. Pass tier through to outpaint options in both the HTTP
route and the pipeline/batch registry. Fix log message to say
"Starting AI canvas expand" and include tier in structured log fields.
Add TIER_PARAMS dict with fast/balanced/high presets controlling band
size, mask dilation, seam strip width, and Telea pre-inpainting. Parse
tier from sys.argv[7] with balanced fallback. Conditional Telea and
seam refinement steps skip cleanly for fast tier. Progressive outpaint
now accepts band_size and progress bounds for tier-appropriate scaling.
- Fix dispatcher pipe deadlock: drain stdout pipe in a background thread
to prevent blocking when ONNX runtime output exceeds 64KB pipe buffer
- Add 5-minute SSE stall timeout so the UI shows an error instead of
hanging forever when async AI processing stalls
- Guard CPU colorization: skip for images >2MP on CPU and when DDColor
model is not installed, with clear user-facing messages
- Add AVIF decode fallback via ImageMagick for bitstream variants that
Sharp's bundled libheif cannot decode (affects all tools)
EXR: add ffmpeg fallback when ImageMagick lacks the OpenEXR delegate
(common on macOS Homebrew installs). HDR: force 8-bit depth output to
prevent CLAHE crash (hist_local requires VIPS_FORMAT_UCHAR). Batch:
disable socket timeout and increase server requestTimeout to 30 min
so large AI batches don't get killed by Node.js defaults.
AVIF (and other Sharp-native formats) were written as raw bytes to a
.png temp file, causing PIL to fail with "cannot identify image file".
Every other AI module wrapper already converts via sharp().png().toBuffer()
before writing; face-landmarks was the only one that skipped this step.
HDR and EXR files decoded by ImageMagick can produce 16-bit PNG buffers.
Sharp's CLAHE operation (hist_local) requires VIPS_FORMAT_UCHAR (8-bit).
Check the buffer depth and convert to 8-bit sRGB before processing.
The wrapper used inline-block for tools without overlay children (e.g.
border). CSS maxHeight: 100% on the img inside an inline-block parent
without explicit height does not resolve, so tall images rendered at
natural height and got clipped by overflow: hidden. Unify both code
paths to always use inline-flex + column, where the flex item properly
shrinks within the container's maxHeight constraint.
Sharp's TIFF encoder silently strips the alpha channel, flattening
transparency against black. This caused border (corner radius, shadow),
beautify, and replace-color tools to produce wrong output for TIFF
inputs when the operation needs transparency. Remove TIFF from
ALPHA_FORMATS so these tools fall back to PNG output.
The find-duplicates tool failed entirely when any uploaded file couldn't
be processed, returning "Duplicate detection failed" or a format-specific
error that aborted the whole batch. With mixed-format uploads (77 files),
this made the tool unusable.
- Skip unprocessable files instead of aborting; return skippedFiles in response
- Switch from fetch() to XHR with upload progress tracking (Uploading X%)
- Add Vite proxy timeout config (5min) to prevent connection drops on large uploads
- Add "Download Grouped" button: ZIP with each duplicate group in its own folder
- Add collapsible skipped-files section in the results UI
- Add 3 integration tests for skip behavior (43 total)
The image element used flex: 1 1 0 (basis 0) inside the inline-flex
preview wrapper, causing it to collapse to zero height whenever a
device frame or watermark overlay was present. Changed to flex: 0 1 auto
so the image uses its intrinsic height and can shrink if needed.
Added regression e2e tests verifying every preset keeps the image
visible, and that adding a watermark does not collapse the image
regardless of whether a frame is active.
setError() in the file store unconditionally set processing to false,
even when called with null to clear a previous error. This meant
setProcessing(true) followed by setError(null) at the start of
processing would immediately kill the processing state, preventing the
progress bar from ever rendering.
Fix setError to only reset processing when there is an actual error.
This also fixes the same latent bug in 10+ other custom tool settings
components that call setProcessing before setError.
Custom-route tools (split, compare, collage, find-duplicates, etc.) only
handled HEIC via ensureSharpCompat, failing on BMP, PSD, RAW, TGA, EXR,
HDR, JXL, and other formats Sharp cannot decode natively. Added the full
decode pipeline from createToolRoute to all 16 affected routes: format
validation via validateImageBuffer, HEIC decoding with actionable error
messages, CLI-based exotic format decoding with nested fallback, and SVG
sanitization to prevent XXE/SSRF/script injection.
The vectorize tool had a custom processing flow that bypassed the
standard useToolProcessor hook -- no progress indication, no server-side
batch, and the Download All ZIP relied on a client-side sequential loop.
Backend: extract core logic into vectorizeBuffer(), register via
registerToolProcessFn() so the /batch endpoint works with p-queue
concurrency and SSE progress events.
Frontend: replace custom fetch loop with useToolProcessor hook and
ProgressCard, giving upload progress, per-file batch status, and
automatic Download All ZIP via the existing tool-page infrastructure.
Also set image/svg+xml MIME type on SVG blobs during batch ZIP
extraction to ensure reliable rendering in <img> tags across browsers.
Adds fileFilter and acceptDescription props to Dropzone so tools can
restrict accepted file types across all entry paths (file picker,
drag-and-drop, paste, single URL import, bulk URL import).
Background images, device frames, custom shadows, and watermark text
were not rendering in the right-pane preview. The preview now updates
in real time for all settings: gradient/solid/image backgrounds, macOS/
Windows/Browser frame chrome, iPhone/MacBook/iPad frame indicators,
custom shadow parameters, and watermark text overlay.
Also fixes a React StrictMode effect-ordering race where the parent
tool-page reset cleared preview state set by the child Settings
component on initial mount.
The stitch route only decoded HEIC via ensureSharpCompat, so exotic
formats (DNG, PSD, TGA, BMP, JXL, HDR, QOI, DDS, ICO, JP2, DPX, etc.)
crashed Sharp at metadata read time, causing "Stitch creation failed".
Add the full CLI decode pipeline (decodeToSharpCompat) matching the
tool-factory pattern, plus SVG sanitization with proper error handling
for each format category.
Replace fetch() with XMLHttpRequest in the frontend to surface upload
progress via onprogress events, and add a progress bar that shows upload
percentage then pulses during server-side stitching.
Replace fragile Unicode-range regex with positive ASCII check
(/^[\x20-\x7E]+$/) that also catches emoji and supplementary plane
characters. Update OpenAPI spec to document percent-encoding.
The X-File-Results header contained raw JSON with non-ASCII characters
from filenames (Chinese, Japanese, etc.), violating RFC 7230. Node.js
threw ERR_INVALID_CHAR on writeHead(). Fixed by wrapping the JSON in
encodeURIComponent() on the backend and decodeURIComponent() on the
frontend, ensuring only ASCII goes into the header while preserving
the original filenames after decoding.
Closes#133
The preview endpoint now returns X-Original-Width/Height headers with
dimensions read from Sharp metadata (or ExifTool for RAW files). The
frontend stores these in FileEntry and the ImageViewer prefers them over
the browser's naturalWidth/naturalHeight, which reflects the resized
preview rather than the original sensor dimensions.
Sharp can read TIFF-based RAW files (DNG, CR2, NEF) directly for
metadata without requiring ImageMagick/darktable to fully decode them.
Try Sharp on the raw buffer first; only fall back to the decode pipeline
for formats Sharp cannot open (PSD, ICO, TGA). For RAW files, enrich
with ExifTool to get real sensor dimensions instead of thumbnail size.
SnapOtter