Adds an opt-in High Quality mode to the Object Eraser, backed by a new inpaint-hq feature bundle (Stable Diffusion 1.5 inpainting via diffusers). The default fast LaMa path is unchanged. Both arch archives are published to deepsafe/feature-bundles and the manifest carries their real sha256/sizes.
Verified end to end: a fresh container pulls the bundle from HuggingFace, checksum-verifies it, extracts torch/diffusers plus the fp16 model, and the HQ sidecar erases a large object with a plausible fill.
Refs #141
removeBackground failures wrap in a SafeError so the specific reason survives the Sentry scrubber; the OOM lighter-model fallback and bridge SafeError passthrough are preserved.
Route both Python exit paths through pythonExitError so the reason survives the scrubber; OOM/segfault stay operational and keep "out of memory" for the lighter-model fallback.
Adds a dedicated remove-gif-background AI tool: removes the background from an animated GIF, WebP, or APNG frame by frame and reassembles a transparent (or composited) animation in WebP, APNG, or GIF, with full per-frame effects. Reuses the background-removal bundle. Verified end-to-end with the real rembg model.
Closes#496.
Follow-ups to the v2.1.0 Sentry telemetry overhaul, found by reviewing live release:2.1.0 events:
- error_code tag was empty because reportError read only the top-level err.code; add extractErrorCode() to walk the cause chain (pg SQLSTATE, node E-code, else first short code).
- InputValidationError from a tool's processV2 in the worker was logged as error_class=bug; classify it as expected for any source. Worker-side ZodError stays a bug (schema drift).
- AI dispatcher timeouts rejected with a bare Error, which the sanitizer scrubbed to a message-less "Error: Error"; reject with an operational SafeError (code "timeout") at both timeout sites.
Each fix written failing-test-first; affected and adjacent unit suites green plus full CI (integration + e2e).
Removes Sentry tracing entirely (BullMQ idle polling burned 4.8M transactions in 2 days at the baked 0.1 rate), decouples PostHog sampling, and replaces the type-only error scrub with a vetted-field sanitizer plus SafeError/ToolInputError contracts. One classified capture path with per-signature throttles and a per-process ceiling makes storms impossible (NODE-1E was 4,541 events from one 30s loop). Browser errors move to a dedicated web Sentry project with their own source maps. Adds the SNAPOTTER_TELEMETRY runtime kill switch and silences test fleets.
Crash fixes: remote 204/304 SSRF process kill (NODE-20), conversion-preset boot crash loop (NODE-21), Redis version preflight + unhandled subscribe rejection (NODE-1T), Sign PDF on plain-http origins (NODE-1K/1M), wavesurfer/pdf.js teardown rejections (NODE-1P/1N), bundle-import ZlibError to 400 (NODE-1Z), chart-maker input errors declassified (NODE-1H/1J), asset requests skip the session DB lookup (NODE-1D).
Make on-demand AI feature-bundle installs reliable and self-healing, closing
the failure modes behind most "some tool doesn't work" reports.
Multi-bundle installs: tools needing more than one bundle (Passport Photo,
Enhance Faces) install every required bundle from one action and stay
not-installed until all are present. Verified across all 19 AI tools.
Downloads: self-heal the accelerated Hugging Face (Xet) client so an upgraded
venv no longer silently falls back to slow urllib; restart instead of
corrupting a resumed partial when a proxy ignores Range and returns 200;
verify the completed size; fail fast on disk-full and HTTP 4xx; retry
transient errors five times; add hf_transfer fallback and document Xet egress.
Install integrity: crash-atomic venv writes so a killed or out-of-space
install can no longer tear the shared venv and break other tools; a boot
breadcrumb reseeds a torn venv to a clean state automatically; a post-install
smoke import test refuses to record a bundle whose libraries cannot load; an
install watchdog stops a wedged installer that would otherwise hold the venv
writer lock forever.
Adds unit and end-to-end tests for every failure mode above.
* fix(api): prevent a crash when an over-limit upload stream has no consumer yet
busboy's "limit" handler destroyed the file stream with an error but never
attached its own error listener, relying entirely on whatever consumes
part.file downstream to do so. On a fast enough connection (or a fully
buffered body, e.g. Fastify inject()), busboy can process enough bytes to
hit the size limit before the route handler's receiveUpload() call has
attached its own stream listener, leaving the resulting "error" event with
zero listeners -- which crashes the whole process by default in Node.
Surfaced by tonight's FULL_MATRIX+FUZZ integration run (880 uncaught
exceptions, all the same root cause). Reproduces deterministically in
isolation; unrelated to this release's actual code delta (file untouched
since PR #413, well before the baseline QA pass).
Fix: attach a baseline no-op error listener the moment the stream is
created, guaranteeing at least one listener always exists. EventEmitter
delivers "error" to every registered listener, so the real consumer's own
error handling is unaffected.
* fix(ai-bundles): rebuild upscale-enhance and photo-restoration to reconcile scipy ABI
upscale-enhance and photo-restoration both depend on codeformer-pip, whose
transitive closure (basicsr -> realesrgan -> gfpgan) pulls in an unpinned
scipy. Both bundles were last built ~June 18-19, before PR #437 added the
manifest's `constraints` array (numpy==1.26.4, scipy==1.12.0, etc.) to pin
exactly this kind of dependency during bundle builds. Only the ocr bundle
was rebuilt after that fix landed.
install_feature.py has no pip install step -- it's a raw tarfile extraction
with no cross-bundle conflict resolution, so installing OCR alongside either
stale bundle left three incompatible scipy versions' files mixed in the same
site-packages directory (a compiled _rotation.*.so from one release next to
Python files expecting a different release's API), breaking the `upscale`
tool and OCR's higher-quality tiers with an ImportError.
Rebuilt both bundles for amd64-gpu and arm64-cpu from the current manifest,
verified scipy/scikit-learn/scikit-image/pandas all resolve to the pinned
versions in the tarballs themselves, then verified end-to-end on real
hardware (Mac arm64 CPU and ubuntu_gpu .248 RTX 4070): installing all
affected bundles together now yields exactly one version of each constrained
package, `upscale` produces correct output, and OCR's balanced/best tiers
correctly use PaddleOCR-GPU instead of erroring out.
Published the rebuilt tarballs to the public deepsafe/feature-bundles
HuggingFace repo and updated this manifest's sha256/sizes to match.
Also adds verify-bundle-compatibility.sh: verify-bundle.sh checks each
bundle in isolation (a fresh venv per bundle), which is exactly why this
shipped twice -- nothing ever checked that bundles built at different times
agree once layered into the one shared venv real installs use. The new
script installs every bundle for an arch into one venv and asserts each
constrained package has exactly one, correct version.
Known follow-up (not fixed here, needs separate discussion): uninstalling a
bundle only removes its downloaded model weights, never the site-packages
it added, so existing installations that already hit this bug have no clean
self-service fix via uninstall+reinstall -- they need a full AI-venv wipe.
* fix(docker): bake a real rate limit default for the all-in-one one-liner
The documented single-container `docker run` install had RATE_LIMIT_PER_MIN=0
(effectively unlimited, ~50k/min) baked in, since only docker-compose.yml
carried a hardened override. A self-hoster following the one-liner path got
no meaningful throttling anywhere, including auth-adjacent routes with no
dedicated per-route limit. Bakes a generous-but-real 1000/min default into
the Dockerfile, raises both compose files' fallback to match so the two
documented install paths converge on the same posture, and updates the Zod
schema default plus docs that quoted the old value.
* fix(api): boot log undercounted tool routes by the conversion-preset total
The "Tool routes: N active" line logged before registerConversionPresets(app)
ran, so it only ever reported the base 158 tools, 83 short of the real
241-tool total. Presets have to register after the base loop (they delegate
to each base tool's own processV2), so the fix moves the log line to after
that call and has registerConversionPresets return its count instead of
reordering the dependency.
* fix(ai): forward {info}/{warning} stderr JSON instead of dropping it
The dispatcher stderr parser only recognized {ready} and {progress,stage}
shaped JSON lines; anything else that parsed as valid JSON (like ocr.py's
GPU-to-tesseract downgrade notice, an {"info": ...} line) matched neither
branch and fell through silently, never reaching docker logs. Adds explicit
{info}/{warning} handling that forwards to console.log/console.warn, same as
the existing [prefix]-tagged non-JSON path.
* fix(api): fall back to a lower OCR tier when PaddleOCR itself is unusable
ocr.ts already retries lower quality tiers on a crashed dispatcher, but the
condition only matched crash-style messages (segfault, exited unexpectedly).
ocr.py's own ImportError/exception handlers already produce messages telling
the caller to use a lower tier (e.g. on the scipy ABI conflict class of bug),
but nothing ever acted on them, so a broken PaddleOCR hard-failed with 422
instead of degrading to Tesseract like ocr-pdf effectively does. Broadens the
retry condition to also catch PaddleOCR-engine-unusable messages.
Note: ocr-pdf's tesseract-only behavior turned out to be an unrelated,
pre-existing, deliberate design choice (PaddleOCR segfaults on rasterized PDF
pages on arm64), not a graceful-fallback mechanism to copy -- the two tools
weren't actually solving the same problem, so this fixes ocr.ts's own gap
rather than trying to mirror ocr-pdf.
* fix: remove all automatic third-party egress (OSM tiles, Scalar fonts, editor Google Fonts, AI model download fallbacks)
Phone-home audit follow-up. The product no longer makes any automatic
third-party request; user-initiated click-outs stay, and production now
fails closed on missing AI models.
1. GPS leak via OSM tiles: the strip-metadata panel auto-loaded
tile.openstreetmap.org tiles encoding the photo's GPS position. The
Leaflet mini-map is gone; coordinates render as text plus an explicit
View on map link (openstreetmap.org, opens on click only). Removed
tile.openstreetmap.org from the CSP img-src, dropped the leaflet
dependency, added the viewOnMap i18n key to all 21 locales.
2. Scalar docs fonts: /api/docs loaded Inter and JetBrains Mono from
fonts.scalar.com. Scalar now renders with withDefaultFonts: false and
both --scalar-font and --scalar-font-code pinned to system stacks;
fonts.scalar.com removed from the docs CSP font-src. Verified by
injecting GET /api/docs/: config carries withDefaultFonts false and
the served page has no fonts.scalar.com reference.
3. Editor Google Fonts: the editor font picker built
fonts.googleapis.com stylesheet URLs for 25 web fonts the served CSP
already blocked. The remote loading path is deleted; the picker now
offers system fonts only, with a SELF_HOSTED_FONTS seam (FontFace API,
same origin) for bundling fonts later. Unknown families saved in old
documents fall back to the browser default.
4. Python sidecar fails closed on model downloads: new
packages/ai/python/offline_guard.py gates every runtime download
fallback (inpaint, outpaint, restore, noise_removal, detect_faces,
enhance_faces, face_landmarks, red_eye_removal, remove_bg, ocr,
transcribe, upscale) behind SNAPOTTER_ALLOW_MODEL_DOWNLOAD=1 with an
actionable error. Bundled models keep working untouched.
5. OCR and transcription library-internal downloads: unbundled PaddleOCR
language and detection fallbacks now raise the guard error naming the
language instead of resolving models over the network; faster-whisper
gets local_files_only when downloads are off.
6. GFPGAN and CodeFormer cwd-relative weights: facexlib and
codeformer-pip resolve helper weights relative to the process cwd and
fetch them from GitHub when absent. They are now symlinked from the
installed bundle files under MODELS_PATH/gfpgan/facelib before the
libraries load, failing closed when unresolvable.
Defense in depth: HF_HUB_OFFLINE=1 and TRANSFORMERS_OFFLINE=1 are set in
the runtime image and in the sidecar spawn env; install_feature.py lifts
them for user-initiated bundle installs and restores them afterwards
(it can run in-process inside the dispatcher). SNAPOTTER_ALLOW_MODEL_DOWNLOAD
is documented in .env.example, default off.
Validation: typecheck 9/9 workspaces, Biome clean on touched files,
5178 unit tests pass, py_compile on all touched scripts, guard behavior
exercised in both dispatcher exec and per-request import modes, zero
remaining runtime references to the three hosts. Docker build and live
AI inference need post-merge verification on the GPU host.
Claude-Session: https://claude.ai/code/session_01XGB4pGvTvb7sUX4JN745U7
* fix: allow AI model downloads by default, make strict offline mode opt-in
Product call: ease of use first. The download gating from the previous
commit inverts its default: runtime model fetches (public model weights
only, never user data) are allowed out of the box so AI tools self-heal,
and SNAPOTTER_ALLOW_MODEL_DOWNLOAD=0 becomes the explicit strict offline
mode for airgapped deployments, where every fallback raises the
actionable error instead of fetching.
Changes: offline_guard blocks only on an explicit 0/false; the
unconditional HF_HUB_OFFLINE/TRANSFORMERS_OFFLINE image ENV is removed
and bridge.ts sets those flags for the sidecar only in strict mode;
.env.example documents the new default; install_feature's lift/restore
stays. All bundled-path preferences, pre-existence checks, and symlink
pre-placement remain, so installed bundles never trigger a download.
The OSM, Scalar font, and editor font fixes are unchanged.
Validation rerun: typecheck 9/9, Biome clean on touched files, 5178
unit tests pass, py_compile on touched scripts, guard behavior verified
for unset/1 (allowed) and 0/false (blocked with the new message).
Claude-Session: https://claude.ai/code/session_01XGB4pGvTvb7sUX4JN745U7
Fixes 15 defects found by a max-effort multi-agent review of the last 6
merged PRs (#388, #390, #391, #392, #393, #394), all adversarially
verified before fixing.
Install queue + dispatcher (the serious cluster):
- features.ts: finalize the installer child exactly once. A failed spawn
fires both "error" and "close", and the second event released the file
lock and active slot that pump() had just handed to the next queued
bundle, letting two pip processes write the same venv concurrently.
Outcome recording now happens before pump() so the next bundle's first
progress frame cannot race the previous install's bookkeeping.
- feature-status.ts: keep failed-install errors in a per-bundle map
instead of the single progress slot. With the queue auto-starting the
next install, the slot was overwritten within seconds and a failed
install vanished without ever surfacing to GET /features.
- bridge.ts: scope child lifecycle per process (stopped-children set +
request generation tags) instead of an instance-wide shuttingDown flag
that the next spawn reset. A stale SIGTERMed child's late close event
could record a phantom crash (5 of which permanently disable the
dispatcher), null out the freshly spawned child, and reject the new
child's pending requests. The request-timeout kill path still counts
as a real crash.
- install_feature.py: the pre-write disk re-check measured ai_dir's
filesystem even when budgeting the cross-filesystem copy that lands on
the venv's disk; now each budget is checked against the filesystem the
bytes actually land on, so ENOSPC cannot strike mid-write and leave
site-packages half overwritten.
Behavior regressions:
- embed-subtitles: preserve pre-existing subtitle tracks (0:s?) and MKV
attachments (0:t?) that the -map 0:v:0/0:a? rewrite silently dropped;
data streams stay unmapped on purpose (the actual MPEG remux fix). The
new subtitle maps first so the language tag hits the right stream.
- usage-survey-overlay: fail closed when the settings fetch fails; the
fail-open path rendered the blocking survey against an unhealthy API
and soft-locked admins, the lock-out class #392 fixed.
- features-store: queued bundles poll instead of each holding an SSE
connection (Install All could pin 7 EventSources and exhaust the
browser's 6-per-origin HTTP/1.1 limit, hanging the whole app);
listenToProgress closes any prior stream and stops any poll before
subscribing; installAll skips bundles already installing or queued.
Contracts, tests, i18n:
- openapi.yaml: add "queued" to the features status enum and document
downloadBytes/installedBytes (Schemathesis conformance).
- feature-lifecycle e2e: queue transcription (~0.5 GB) instead of ocr
(~6 GB) and give the test a budget that covers both install drains
(the stacked waits exceeded the old 900s timeout).
- docker-compose.qa.yml: parameterize the host port (QA_APP_PORT) so
QA_PROJECT_NAME concurrent stacks can actually bind.
- compare + watermark-image: restore per-input error attribution
("Invalid first/second image", "Invalid watermark image") lost in the
shared-handler migration.
- ai-features-section: the "{size} on disk" suffix now goes through
i18n; key added to all 21 locales.
- watermark-image + content-aware-resize: migrate to the shared
inputHandlerFor("image") chain like compare/vectorize/compose, fixing
drift in the inline copies (no SVG sanitize, no RAW extension hint,
no AVIF probe).
Verified: typecheck across 9 workspaces, Biome clean on all changed
files, 584 targeted unit tests and 249 integration tests green
(including real-ffmpeg embed-subtitles runs). One unit test updated to
the new poll-while-queued contract with a single-EventSource assertion.
Claude-Session: https://claude.ai/code/session_017mR1HiHaf3a1BmUtrHX4j3
* fix(api): correct format/filename/container handling across tool routes
Found during a comprehensive QA sweep exercising every tool against its
full accepted-format matrix:
- watermark-image, compose: preserve the requested output format and a
matching download filename/extension instead of always emitting the
source format
- compose: crop oversized overlays to the visible base area instead of
crashing Sharp's composite, and reject only overlays fully outside the
base image instead of any oversized one
- compare, vectorize: switch to the shared image input handler so
filenames and formats like .svgz/.tga/RAW survive validation instead
of being rejected pre-processing
- tool-factory, images-to-video: normalize frames through Sharp before
handing them to FFmpeg, fixing GIF/AVIF/RAW image-to-video jobs that
previously failed or hung
- media-tool, replace-audio, embed-subtitles: fix legacy container
MIME/codec handling for MPEG sources and subtitle remux cases
- files: expand download MIME mapping for text/data/document/video/audio
outputs that were falling back to a generic content type
- convert-document/presentation/spreadsheet: same-format conversions now
return the original validated file instead of erroring or producing
corrupt tiny output
Claude-Session: https://claude.ai/code/session_019fpSXhLGLXWwfyZY2tWhLG
* fix(web): dropzone a11y, stale localStorage getter, dead code
- dropzone: stop making the whole drop-zone section clickable/focusable.
A section acting as an interactive element around a real upload button
is a nested-interactive-element anti-pattern that confuses screen
readers; drag-and-drop doesn't need focus semantics, only the button
fallback does. Keeps that button semantic and keyboard-reachable.
Updates the two e2e call sites that clicked the section directly.
- api, use-auth: read through window.localStorage via the existing API
storage helper instead of the bare global, which resolves to Node's
experimental localStorage getter under Vitest and threw
- find-duplicates-settings, info-settings, login-page: remove dead code
(unused zip-download handler, a stale mount-only effect dependency
that left cached info stuck at reused indices, an unused response
variable)
Claude-Session: https://claude.ai/code/session_019fpSXhLGLXWwfyZY2tWhLG
* fix(i18n): pt-BR, zh-CN, zh-TW were silently falling back to English
The locale loader looked up dynamic-import exports by the raw locale
code (mod["pt-BR"], mod["zh-CN"], mod["zh-TW"]), but those three modules
export camelCased bindings (ptBR, zhCN, zhTW) since identifiers can't
contain hyphens. The lookup returned undefined and every consumer
silently fell back to English for these three locales. Replaces the
generic lookup with explicit per-locale loaders so the mapping can't
drift out of sync again.
Also updates the dropzone helper copy across all 21 locales to match
the drag-only dropzone wording from the previous commit.
Claude-Session: https://claude.ai/code/session_019fpSXhLGLXWwfyZY2tWhLG
* fix(docs): clear build warnings in the VitePress site
- config.mts: add an onwarn handler for the @vueuse INVALID_ANNOTATION
warnings emitted during the docs build
- deployment.md: the caddyfile code fence language isn't a shiki grammar
VitePress ships with, so it warned on every build; use txt instead
Claude-Session: https://claude.ai/code/session_019fpSXhLGLXWwfyZY2tWhLG
* test(qa): update QA harness for the drag-only dropzone and regen metadata
- api-sweep, qa-helpers, verify-ai: add JSON-body tools, multi-input
secondary fixtures, async polling for slow valid jobs, 501
FEATURE_NOT_INSTALLED skip handling, and safer per-tool settings
- input-preview, pipeline-ui specs: update upload flow for the
drag-only dropzone surface
- add tests/fixtures/data/valid/chart.json, a valid chart fixture the
updated helpers route to
- regenerate tools-meta.json against current TOOLS[]
Claude-Session: https://claude.ai/code/session_019fpSXhLGLXWwfyZY2tWhLG
* fix(security): close a login timing side-channel, harden zip-slip tests
Found during a black-box security sweep of the real auth-enabled
production container: a nonexistent username returned 401 in ~3-10ms,
while a wrong password for a real user took ~35-42ms, because scrypt
verification only ran when a user row existed. That timing gap lets an
attacker enumerate valid usernames without ever guessing a password.
Now runs verification against a cached dummy hash on the unknown-user
path too, so both cases cost the same regardless of outcome.
extract-zip already had a relative-traversal regression test
(../evil.txt), but its absolute-path rejection branches
(name.startsWith("/") / startsWith("\\")) had none. Added the three
missing cases: deep relative traversal, absolute Unix path, and
Windows-style absolute path.
Claude-Session: https://claude.ai/code/session_019fpSXhLGLXWwfyZY2tWhLG
* test(qa): add UI-driven AI bundle install scripts
QA_PROMPT.md's Phase 2 requires installing AI models the way a user
does -- through the UI, on demand from HuggingFace -- and treats the
curl-based admin install endpoint as fallback-only. Nothing in the
harness actually drove that flow; tests/qa/seed-ai-models.sh installs
via docker exec + pip, which is further from a real user than even the
API fallback.
install-ai-bundles-ui.mts logs in, opens Settings > AI Features,
screenshots the pre-install state, clicks Install All, and screenshots
progress -- then exits, since installs continue server-side once
triggered. verify-ai-install-complete.mts polls bundle status,
screenshots the completed state, and runs one real tool per installed
bundle to prove the freshly-downloaded model actually executes.
Claude-Session: https://claude.ai/code/session_019fpSXhLGLXWwfyZY2tWhLG
* fix(qa): correct the apiToolPath import in the AI verify script
Dynamic import of the package name failed under tsx's module resolution
from apps/api's node_modules context; use the same relative-path import
api-sweep.mts already uses successfully.
Claude-Session: https://claude.ai/code/session_019fpSXhLGLXWwfyZY2tWhLG
* fix(web): correct AI bundle size estimates shown before install
Measured real downloads during GPU-node QA verification: photo-restoration
pulls ~4.4GB (was advertised as 800MB-1GB, off by 4-5x) and ocr pulls
~5.5GB (was advertised as 3-4GB). Both estimates only accounted for model
weights, not the pip dependencies (torch/paddle) that come down with them.
Updated to reflect actual total download size, since that's what a user
deciding whether they have the disk/bandwidth actually needs to know.
Claude-Session: https://claude.ai/code/session_019fpSXhLGLXWwfyZY2tWhLG
* fix(web): make desktop Settings reachable when auth is disabled
AvatarDropdown (the only desktop entry point to Settings) was gated
behind `!isMobile && authEnabled`. With AUTH_ENABLED=false the synthetic
anonymous admin user should have full Settings access per how auth.ts
documents this mode -- and the mobile bottom nav already worked this way,
showing Settings unconditionally. Desktop just had a stray extra gate the
component doesn't need: AvatarDropdown already resolves its own username
internally (falling back to "admin") and reads authEnabled itself where
it actually matters (hiding the Logout button). Removed the outer gate;
verified end-to-end against a fresh AUTH_ENABLED=false instance -- avatar
now renders, Settings opens, shows the anonymous/Admin identity correctly.
Also documents (not changes) a related finding in install_feature.py:
detect_arch() always resolves amd64 hosts to the GPU-bundled archive
variant regardless of actual GPU presence, since no CPU-only amd64
archive is published to the bundle repo yet. Left as a code comment
rather than a behavior change, since requesting an unpublished archive
key would hard-fail installs entirely -- worse than the current
oversized-but-working download. Full detail in the QA report.
Claude-Session: https://claude.ai/code/session_019fpSXhLGLXWwfyZY2tWhLG
* fix(ai): stop logging expected dispatcher reloads as crashes
After each AI bundle install the Python dispatcher reloads because the
venv changed, and after every app shutdown it's SIGTERMed. Both took the
close handler's `code !== 0` branch (SIGTERM makes the exit code null),
so they were counted as crashes -- producing an alarming "crash" line in
the logs and a pointless ~1s recovery backoff after each of 7 installs.
A `stopping` flag set in shutdown() lets the close handler tell an
intentional stop apart from a real crash. The request-timeout kill path
deliberately does not set it, so a genuinely hung script still records a
crash and the 5-in-60s permanent-disable threshold is untouched.
Claude-Session: https://claude.ai/code/session_019fpSXhLGLXWwfyZY2tWhLG
* fix(api): return a clean message when content-aware resize times out
Carving a very high-resolution image down to a tiny target could exceed
the caire subprocess timeout, and the raw error forwarded to the user was
caire's terminal output -- ANSI color codes and progress-spinner control
characters -- instead of anything actionable. Now: the timeout path
throws a clear "timed out; try a smaller image or larger target" message
(keeping the raw stderr as `cause` for server logs); friendlyError()
strips ANSI/control chars centrally so any subprocess dump surfaced
through the shared sanitizer is plain text; and the content-aware-resize
route (a custom route that bypassed the sanitizer) now routes its error
paths through friendlyError like every other tool.
Claude-Session: https://claude.ai/code/session_019fpSXhLGLXWwfyZY2tWhLG
* fix(ai): stop bundle installs from exhausting host disk
Installing an AI bundle on a tight-disk host could push the root
filesystem to zero bytes free after the preflight check had already
passed. Two root causes:
- move_tree used copytree+rmtree, so during the move the extracted
payload existed in both staging and the venv at once -- a full
transient doubling on disk. Rewrote it to rename entries (a cheap
metadata op on the same filesystem, no copy), falling back to a copy
only across filesystems.
- the preflight budget used the manifest's extractedSize verbatim, which
is 0 for several archives, collapsing the estimate to just the
compressed size. Added a conservative fallback (3x compressed) so a
missing value can't under-reserve.
Also added a real-on-disk re-check immediately before the first
destructive venv write (measuring the actual extracted payload and
whether the move needs extra space for a cross-filesystem copy), which
also now covers the offline-import path that previously skipped the disk
check entirely; wrapped the moves so an out-of-space failure returns a
clean actionable error instead of a traceback; and made the disk check
resolve the nearest existing ancestor so it never throws on a
not-yet-created venv path.
Claude-Session: https://claude.ai/code/session_019fpSXhLGLXWwfyZY2tWhLG
* feat(web): show the real per-arch AI bundle download size
The bundle cards and install prompt showed a hardcoded, architecture-blind
estimatedSize string. That's misleading: amd64 hosts always pull the
CUDA-inclusive archive (there's no CPU-only amd64 variant published), so a
bundle labelled "1-2 GB" can actually download several times that, while
arm64 pulls a much smaller archive for the same label. The manifest
already carries the real per-arch compressedSize (and extractedSize where
measured), so surface those: a new optional downloadBytes/installedBytes
on FeatureBundleState, populated in getFeatureStates() for this host's
arch (resolver mirrors install_feature.py detect_arch), shown by the UI
when present with estimatedSize kept as the fallback label. Also nudged
upscale-enhance's fallback string (4-5 -> 5-6 GB) to match its real
compressed size, consistent with the earlier photo-restoration/ocr fixes.
Fields are optional so demo/mock and existing tests stay compiling; the
manifest's extractedSize is 0 for a few archives, which now surfaces as
null rather than a bogus 0.
Claude-Session: https://claude.ai/code/session_019fpSXhLGLXWwfyZY2tWhLG
* fix(web): move the AI install queue to the server so it survives tab close
Installing multiple bundles could silently lose all but the first. The
server rejected a concurrent install with 409, so the client worked
around it by queueing the rest in browser-local state and only POSTing
each once it saw the previous finish. A single POSTed install is durable
(the installer child is detached from the request), but a queued one had
zero server footprint -- close the tab mid-queue and those installs
vanished with no error, while the UI still showed them "Queued". The
client "mutex" didn't even serialize: the queued bundles' local waits all
resolved at once and raced into concurrent POSTs that 409'd each other.
Now the queue lives on the server (a small in-memory FIFO leaf module).
The install endpoint enqueues instead of 409-ing and returns
202 {jobId, queued}; a pump starts the next bundle when the current one's
child exits (and after an offline import releases the lock), all behind
the existing venv + file locks, which are unchanged. The client just
POSTs every bundle immediately and reflects the server-reported
queued/installing status; Install All fires all POSTs and lets the server
serialize them, keeping the one-shot retry-on-failure. Adds "queued" to
FeatureStatus (the bundle card already rendered that state) and surfaces
it from getFeatureStates. In-memory is deliberate: it matches the
existing contract (survives a tab close, not a server restart, which
already clears the lock on boot).
Claude-Session: https://claude.ai/code/session_019fpSXhLGLXWwfyZY2tWhLG
* fix(qa): don't log env-derived credentials in the AI-install script
CodeQL flagged clear-text logging of sensitive information: the login
status line interpolated the QA base URL and username (both read from
the process environment) into a console.log. Replaced with a static
message. QA helper only, but it's a real hygiene issue and cleared the
high-severity code-scanning alert on the PR.
Claude-Session: https://claude.ai/code/session_019fpSXhLGLXWwfyZY2tWhLG
Found and fixed during a full local Docker build validation (amd64/arm64, all
four fleet targets, AI bundle installs, QA harness) and the follow-up bug
sweep requested afterward. None of the affected scripts run in CI, so these
had been silently broken indefinitely.
- docker/feature-manifest.json: pythonVersion was a flat "3.11", but the
amd64 base (Ubuntu 24.04) ships Python 3.12 while arm64 (Debian bookworm)
ships 3.11. Changed to a per-arch object matching the file's existing
convention.
- tests/qa/api-sweep.mts and verify-ai.mts: bare "@snapotter/shared" import
can't resolve since tests/ is not a pnpm workspace member, making both
silently unrunnable via their own documented command on any fresh
checkout. Switched to a relative import.
- tests/qa/generate-ledger.mts: wrote to docs/qa/ without creating the
directory first; docs/ is gitignored except COMMUNITY_GUIDE.md, so a fresh
checkout threw ENOENT.
- Seven QA Playwright spec files (input-preview, settings,
settings-extended, multifile, output-preview, pipeline-ui, smoke) had
~115 fixture() calls using directory names that don't exist. Resolved
every call programmatically against the real fixture tree.
- packages/ai/src/bridge.ts: AI dispatcher restart (happens on every bundle
install) was falsely counted as a crash, risking permanent dispatcher
disable after enough legitimate restarts within the crash window. Added a
shuttingDown flag checked at all three recordCrash() call sites.
- packages/image-engine/src/operations/auto-enhance.ts: image-enhancement
hung 40+ seconds on large RAW photos (confirmed on a real 20.2MP file) in
Sharp's .clahe() step, whose cost scales with total pixel count regardless
of tile size. Added a 16-megapixel cap above which CLAHE is skipped;
verified against the real file (40+s -> 2.0s) with no regression to other
RAW formats or normal-sized images. Fixing this surfaced a second,
smaller bug where the saturation step's CLAHE compensation boost was
keyed off the raw toggle instead of whether CLAHE actually ran.
- Two QA-harness robustness gaps closed per "fix everything, even the small
bugs": the passport-photo/erase-object input-preview tests now skip
cleanly with a clear reason on a container without their AI bundle
installed, and docker-compose.qa.yml's hardcoded project/container name
(the actual root cause of a mid-validation container swap between two
concurrent sessions) is now parameterized via QA_PROJECT_NAME.
Full validation report is local-only per repo convention.
* fix(docker): pin CUDA base to 12.6 so the GPU image starts on R560+ drivers
The amd64 base nvidia/cuda:12.9.2-cudnn-runtime bakes a cuda>=12.9 driver gate enforced by nvidia-container-toolkit at container start, so the image fails to launch on common production drivers (e.g. 570.x / CUDA 12.8). The AI bundles are all cu126 wheels and the image installs libcublas-12-6, so 12.9 was misaligned with the workload. Pin to nvidia/cuda:12.6.3-cudnn-runtime-ubuntu24.04 to match the wheels and lower the driver floor to R560+.
* fix(ai): broaden OOM detection so the rembg lighter-model fallback fires
onnxruntime/CUDA allocation failures surface as 'Failed to allocate memory for requested buffer', CUBLAS_STATUS_ALLOC_FAILED, or bad_alloc, not just 'out of memory'. The background-removal and transparency-fixer fallback-to-lighter-model paths only matched the literal 'out of memory', so the fallback was dead code and transparency-fixer (default birefnet-hr-matting) always failed with an allocation error. Add isMemoryAllocError() and use it in both checks.
* fix(ai): use bundled PaddleOCR models so OCR runs offline
ocr.py passed no model dirs to PaddleOCR, so PaddleX resolved models from ~/.paddlex and downloaded them from HuggingFace at runtime (slow first use, broken air-gapped), ignoring the models the OCR bundle ships in MODELS_PATH; it also pulled doc-orientation/unwarping models that are not bundled. Pin detection, recognition and textline models to the bundled dirs in MODELS_PATH (per language) and disable use_doc_orientation_classify / use_doc_unwarping, with per-component fallback when a model is absent. Verified: OCR runs with zero HuggingFace requests.
* fix(docker): add CAP_KILL so container shutdown is graceful
cap_drop: ALL without re-adding KILL meant tini (PID 1, root) could not forward SIGTERM to the gosu-dropped snapotter process (root minus CAP_KILL cannot signal a different UID). docker stop logged '[FATAL tini] forwarding signal: Operation not permitted', never delivered the signal, and fell back to SIGKILL after the 10s timeout. Add KILL to cap_add in both compose files. Verified: docker stop completes in 0s with SIGTERM delivered (exit 143) and no FATAL tini.
* fix(ai): serialize bundle installs against AI jobs to prevent sidecar segfault
A feature bundle install rewrites the shared Python venv (pip + copytree of site-packages/*.so) as a background subprocess, with no coordination against AI tool jobs that dlopen native libs (torch / onnxruntime CUDA) from the same venv; a job loading a shared object while it is overwritten segfaults the sidecar. Add a process-wide async mutex (venv-lock.ts): bridge.run() acquires it before every AI script and the install route holds it across the installer subprocess. Both run in the same Node process so a module-level lock suffices. Verified: concurrent install + AI job produces zero segfaults and the job serializes behind the install.
* fix(ai): make the venv lock read/write so concurrent AI jobs are not serialized
The first cut used an exclusive mutex, which (a) deferred the dispatcher spawn by a microtask and broke unit tests that synchronously drive the mocked spawn, and (b) serialized AI jobs against each other, removing the dispatcher's by-id request multiplexing. Make it a writer-preferring read/write lock: AI jobs are shared readers (with a synchronous fast path so spawn still happens in-tick) and a bundle install is the exclusive writer. Verified: all 764 AI unit tests pass.
* fix(ai): degrade OCR to Tesseract on CPU-only hosts instead of segfaulting
The amd64 AI bundle ships paddlepaddle-gpu, whose native libs dlopen
libcuda.so.1 at import and segfault on a host without a GPU (libcuda is the
driver lib, injected only by nvidia-container-toolkit on GPU hosts). The
segfault crashed the shared long-lived AI dispatcher and, after a few attempts,
tripped the bridge crash-recovery permanent-disable, wedging all AI until a
container restart. The standalone ocr tool defaults to quality=balanced
(PaddleOCR), so it hit this on every CPU-only deployment; ocr-pdf already
hardcoded Tesseract and was unaffected.
ocr.py now gates the PaddleOCR tiers on gpu_available(): balanced/best
transparently fall back to fast (Tesseract, CPU-capable) when no usable GPU is
present, and run_paddleocr_v5/run_paddleocr_vl refuse before importing paddle so
the GPU build is never dlopen'd on CPU. GPU hosts are unchanged.
Verified on a CPU-only Windows/WSL2 box: ocr returns Tesseract text across
repeated runs with the dispatcher staying healthy (no wedge).
The persistent Python dispatcher rejects scripts whose feature bundle is not
installed, but the per-request fallback (used when the dispatcher is down, e.g.
restarting right after a model repair) spawned scripts directly and bypassed
that gate. Behavior was therefore inconsistent: a gated script would fail under
the dispatcher but run under the fallback -- the "works once after a repair"
symptom from the original report.
- add packages/ai/src/feature-gate.ts: SCRIPT_BUNDLE_MAP + missingBundleForScript,
mirroring TOOL_BUNDLE_MAP in dispatcher.py, reading the same installed.json and
failing closed exactly like dispatcher._get_installed_bundles()
- runPerRequest now rejects with "feature_not_installed" (the same message the
dispatcher path surfaces) when a gated script's bundle is not installed
- unit tests for the gate, plus a drift test pinning the TS map to dispatcher.py
Closes#327
* feat(tracing): add OpenTelemetry dependencies and --import preload flag
* feat(enterprise): add distributed_tracing feature gate
* feat(tracing): add SDK bootstrap with enterprise gating
* fix(tracing): correct test coverage for enterprise-unavailable path and prevent double-init
Test 2 now mocks @snapotter/enterprise to throw an import error, exercising
the catch block in the preload. Test 3 imports with no endpoint so the preload
is a no-op, avoiding leaked SDK from double-initialization. Added idempotency
guard to initTracing() as a safety net.
* feat(tracing): add Pino trace mixin and shared logger
When OTel tracing is active, every Pino log line now includes traceId,
spanId, and traceFlags fields for log-to-trace correlation. The mixin
is a no-op when no SDK is registered (community users).
* feat(tracing): add _otel to ToolJobData and inject trace context at enqueue
Add optional _otel carrier field to ToolJobData for W3C trace context
propagation across BullMQ job boundaries. When an active OTel span exists,
propagation.inject() writes traceparent/tracestate into the job data before
queue.add(). When no SDK is registered (community edition), the carrier
stays empty and _otel remains undefined -- zero overhead.
* feat(tracing): extract trace context and create spans in BullMQ worker
* feat(tracing): inject trace context into Python sidecar calls
* feat(tracing): add trace context extraction to Python sidecar
* feat(tracing): add shutdownTracing to graceful shutdown sequence
* feat(tracing): enrich HTTP spans with tool_id and user_id attributes
* docs: add OpenTelemetry env var documentation to .env.example
* test(tracing): add lifecycle integration tests for trace propagation
* fix(tracing): inject trace context into pipeline and batch flow jobs
* fix(tracing): add sidecar.execute Node-side span and remove unnecessary comment
Wraps PythonDispatcher.run() with a sidecar.execute span on the Node
side so traces show the full round-trip (Node span -> Python span).
Also removes an obvious comment from logger.ts.
- Expose birefnet-hr-matting in UI (People/Ultra) and fix model defaults
(People/Max now uses birefnet-matting for true alpha matting)
- Add output format selector (PNG/WebP/AVIF) with lossless alpha support
- Add edge smoothing post-processing (Off/Light/Medium/Strong) via
morphological mask refinement to reduce gray halo artifacts
- Add color decontamination to remove background color spill from
semi-transparent edge pixels
- Thread new settings through full stack: frontend -> API schema ->
Python sidecar -> Sharp effects pipeline
- Add i18n keys for all 21 locales
- Add unit tests for new option serialization (3 tests)
- Add integration tests for new settings validation (4 tests)
- Filter known client-error noise (rate limit, empty body, unsupported
media type, content-length mismatch, premature close) from Sentry
via beforeSend to stop 644 events of non-actionable noise
- Sanitize x-output-filename header to prevent TypeError on non-ASCII
filenames in optimize-for-web preview (23 events)
- Handle EPIPE on Python dispatcher stdin write with graceful fallback
to per-request spawning instead of crashing (NODE-W)
- Map EACCES on storage directory/file write to proper 503 status
instead of generic 500 (NODE-P, 3 events)
The GPU detection in gpu.py had two issues preventing GPU usage in
containers (especially rootless podman with CDI):
1. When torch was installed but torch.cuda.is_available() returned
False, the function returned immediately without trying the
ONNX Runtime + nvidia-smi fallback. This meant a CPU-only torch
build (installed before GPU was available) would block all GPU
detection, even for ONNX-based tools.
2. The failure logged a generic "torch loaded but CUDA not available"
with no diagnostic information, making it impossible to debug
whether the issue was a CPU-only build, missing libraries, or
device permissions.
The fix restructures gpu_available() into three detection tiers
(torch -> ONNX Runtime -> nvidia-smi) that always fall through on
failure. When torch CUDA fails, it now checks torch.version.cuda to
distinguish CPU-only builds from CUDA builds that can't access the
GPU, and logs LD_LIBRARY_PATH, torch.cuda.init() errors, and
nvidia-smi results.
Also fixes two env var passthrough bugs in buildMinimalEnv():
- SNAPOTTER_GPU was never passed to the Python subprocess, so the
user-facing GPU override env var had no effect
- MODELS_DIR was a dead entry (never set as env var); replaced with
MODELS_PATH which the Dockerfile sets and Python scripts read
Closes#134
Auth: login rate limit 30/min (was 500), global rate limit 1000/min (was
unlimited), password/username max lengths on all Zod schemas, session
invalidation on role change, API key legacy scan bounded to 100 keys.
SVG: hardened regex sanitizer with CDATA stripping, XML entity decoding,
set/animate/iframe/embed blocking, comprehensive data: URI blocking,
use element external href blocking. 11 attack payload fixtures added.
SSRF: fixed DNS rebinding TOCTOU by pinning resolved IPs via custom
HTTP/HTTPS agents. Added 6to4 and NAT64 to blocked IPv6 ranges.
Docker: capability dropping (cap_drop ALL + minimal cap_add), resource
limits (4g/8g mem, 512/1024 pids), healthcheck timeout, password
removed from startup banner, default password warning comments.
Network: CSP and HSTS applied in all environments (not just production),
stack traces removed from all error responses, internal paths stripped
from error details, per-route rate limits on uploads (60/min) and URL
fetches (200/hour).
Files: exclusive temp file creation (O_EXCL), disk space circuit
breaker, per-user storage quotas, settings payload 64KB size guard.
Python sidecar: script name allowlist in dispatcher, minimal environment
for subprocess spawns.
Dependencies: fixed 6 production CVEs (drizzle-orm, fastify, fast-uri,
@fastify/static, next, archiver/lodash). Pinned all GitHub Actions to
SHA hashes.
114 security tests added. Full OWASP Top 10 penetration test matrix
verified against production Docker container (30/30 pass after
hardening).
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.
- Refactor use-tool-processor and use-pipeline-processor hooks
- Enhance dropzone component with improved UX
- Improve seam carving with better error handling and tests
- Add JXL format encoding support to format-encoders
- Update tool routes for consistent format handling
- Add dropzone unit tests
The upscale function called runPythonWithProgress without a timeout parameter,
defaulting to the bridge's 10-minute hard limit. On CPU-only systems like
Synology NAS devices, Real-ESRGAN 4x upscaling easily exceeds this for modest
images. Additionally, when the timeout fired on the dispatcher path, the Python
process was left running and blocked all subsequent AI operations.
This fix adds an adaptive timeout based on input megapixels, scale factor, and
GPU availability (180s/effective-MP on CPU, 30s/effective-MP on GPU, floor of
10 minutes). It also kills the dispatcher on timeout so subsequent requests can
proceed via a fresh restart.
Closes#119
The dispatcher was lazy-initialized on first AI request, but a race
condition meant the first call always missed it (dispatcherReady still
false) and fell through to cold per-request Python. initDispatcher()
starts the dispatcher eagerly and returns a Promise that resolves with
GPU status once ready (or after a timeout).
The close handler called recordCrash() unconditionally, even for exit
code 0 (normal MAX_REQUESTS restart). After 5 normal cycles within 60s
the dispatcher was permanently disabled. Now only non-zero exits count.
- Add 55 unit tests for feature-status.ts (installed.json CRUD, cache
behavior, install lock, model verification, crash recovery, composite
state) using real temp directories
- Add 36 integration tests for full install/uninstall lifecycle against
Docker containers (face-detection bundle, SSE progress, tool gates,
shared model protection, concurrent install prevention, auth guards,
container restart recovery)
- Fix noise-removal CPU timeout by adding megapixel-based timeout
calculation (120s/MP, min 5 minutes)
- Fix Playwright auth storage state race condition (mkdirSync before
saving analytics-user.json)
- Fix 2 skipped tests in fixes-verification.spec.ts by replacing
external ~/Downloads/sample dependency with existing test fixtures
- Enable skipped analytics-consent settings toggle test
- Restructure features.spec.ts to manage bundle state (uninstall/
reinstall OCR) so 501 guard tests run instead of skipping
- Update noise-removal test mock to include sharp metadata() method
Skip alpha matting on CPU (pymatting's sparse matrices are the main
memory hog), auto-downscale images above 2048px before sending to
rembg, and retry with the lighter u2net model when OOM is detected.
Closes#17, #18, #19, #31, #32, #33, #34
Format preservation (#17, #18, #19):
- Add resolveOutputFormat to rotate, resize, text-overlay, watermark-text,
border, replace-color, blur-faces, upscale, erase-object, restore-photo
- Alpha-aware fallback: border with corner radius/shadow and replace-color
with makeTransparent fall back to PNG for non-alpha formats (JPEG)
- Python sidecar tools (blur-faces, upscale, erase-object) now convert
PNG output back to input format, matching restore-photo/colorize pattern
- Upscale and erase-object default to "auto" format detection instead of PNG
Dispatcher stability (#31, #32):
- Add gc.collect() and torch.cuda.empty_cache() after each dispatcher request
- Add configurable max_requests (default 50) for periodic dispatcher restart
- Add exponential backoff to dispatcher crash recovery in bridge.ts
- Circuit breaker: 5 crashes within 60s permanently disables dispatcher
- Reset crash counter on successful dispatcher startup
Health & security (#33, #34):
- Export getDispatcherStatus() from @snapotter/ai with running/ready/failed/
gpu/pid/consecutiveCrashes fields
- Admin health endpoint now includes full dispatcher status
- Add pip-audit job to CI workflow for Python dependency scanning
Code fixes:
- Sidebar state bleed: reset file store on HomePage mount
- restore-photo: raise error instead of silently skipping colorize
when DDColor model missing
- PaddleOCR OOM: cap input images to 2048px before OCR inference
- Torch CPU optimization: use --index-url .../whl/cpu on CPU nodes
Test fixes:
- upscale: add exact:true to scale factor button locators
- smart-crop: add exact:true to "Pad to square" locator
- colorize: use regex for model button names (Best/Balanced/Fast)
- enhance-faces: use .first() for ambiguous percentage display
- passport-photo: fix DPI locator, .or() compound, generate fallback
- people: update maxUsers assertions for unlimited (0) default
- automate: "Save Pipeline" → "Save" matching actual button text
- tools.test: add resize to Sharp mock chain for OCR tests
- Add enable_mkldnn=False to PaddleOCR constructor to bypass PaddlePaddle
3.3+ OneDNN/PIR crash on CPU-only systems
- Add 25MP and 75% max-reduction guard to seam carving with clear error
messages instead of silent timeout/crash
- Replace barcode/QR AVIF test fixtures with actual scannable codes
(old fixtures did not contain real barcodes)
- Convert all AI bridge inputs to PNG before writing to disk so PIL can
read AVIF/WebP/TIFF (7 bridge files; face-detection and OCR already
had this pattern)
- Add title/author aliases to edit-metadata schema so common field names
actually write EXIF tags instead of being silently stripped by Zod
- Port extend/pad crop logic from passport-photo single endpoint to the
batch pipeline so crop regions extending beyond the image get filled
with background color instead of producing all-white output
- Clamp quantized color channels to 255 in color-palette to prevent
Math.round(255/16)*16=256 from producing invalid hex like #100100100
- Compare OCR fallback warning against expected engine name per tier
instead of comparing engine name against tier name (always mismatch)
- Added model mismatch warnings in colorize, enhance-faces, and upscale routes.
- Improved error handling in colorize, enhance_faces, remove_bg, restore, and upscale scripts with detailed logging.
- Updated Dockerfile to align NCCL versions for compatibility.
- Introduced a new full tool audit script to test all tools for functionality and GPU usage.
- Created Playwright E2E tests for GPU-dependent tools to ensure proper functionality and performance.
- Replace [object Object] errors with readable messages across all 20+ API
routes by normalizing Zod validation errors to strings (formatZodErrors)
- Add parseApiError() on frontend to defensively handle any details type
- Add global Fastify error handler with full stack traces in logs
- Fix image-to-pdf auth: Object.entries(headers) → headers.forEach()
- Fix passport-photo: safeParse + formatZodErrors, safe error extraction
- Fix OCR silent fallbacks: log exception type/message when falling back,
include actual engine used in API response and Docker logs
- Fix split tool: process all uploaded images, combine into ZIP with
subfolders per image
- Fix batch support for blur-faces, strip-metadata, edit-metadata,
vectorize: add processAllFiles branch for multi-file uploads
- Docker: LOG_LEVEL=debug, PYTHONWARNINGS=default for visibility
- Add Playwright e2e tests verifying all fixes against Docker container
Set U2NET_HOME=/opt/models/rembg so rembg models pre-downloaded at
build time as root are found at runtime by the non-root ashim user.
Without this every fresh container re-downloaded the 973 MB BiRefNet
models on first background-removal request.
Apply the same fix to PaddleOCR: download to /opt/models/paddlex and
symlink into both /root/.paddlex and /app/.paddlex so PaddleX finds
models regardless of which HOME gosu resolves at runtime.
Fall back to per-request spawning in bridge.ts when the persistent
dispatcher crashes mid-request (e.g. OOM loading a large ONNX model),
so the operation succeeds instead of surfacing "Python dispatcher
exited unexpectedly" to the user.
Improve entrypoint.sh permission warning to mention Windows bind mounts
as the likely cause.
* feat(passport-photo): add passport specs database and tool constants
* feat(passport-photo): add MediaPipe FaceMesh landmark detection script
* feat(passport-photo): add TypeScript bridge for face landmark detection
* feat(passport-photo): add API routes with analyze and generate endpoints
* fix(passport-photo): accept landmarks from request body and fix pixel coordinate conversion
- Generate endpoint now accepts landmarks + imageWidth/imageHeight in request body
instead of re-running AI face detection (makes generate phase instant)
- Fixed bug where normalized landmark coordinates (0-1) were used directly
as pixel values in crop computation - now properly multiplied by imgW/imgH
- Fixed same bug in pipeline process function
* feat(passport-photo): add UI component with live preview and compliance overlay
---------
Co-authored-by: stirling-image <stirling-image@users.noreply.github.com>
Add comprehensive photo restoration tool that chains multiple AI models:
- Scratch/tear/spot detection via morphological analysis (top-hat/black-hat transforms)
- Damage inpainting via LaMa ONNX model (reuses existing infrastructure)
- Face enhancement via CodeFormer ONNX (~377MB, from facefusion/models-3.0.0)
- Noise reduction via OpenCV NLMeans in LAB color space
- Optional B&W auto-colorization via DDColor (reuses existing model)
Settings: 3 restoration modes (Light/Auto/Heavy), individual feature toggles
for scratch removal, face enhancement (with fidelity slider), denoising
(with strength slider), and auto-colorize. Before/after comparison view.
Handles HEIC, HEIF, and all standard formats. Batch processing supported.
No new Python dependencies - reuses onnxruntime, cv2, mediapipe, PIL.
Co-authored-by: stirling-image <stirling-image@users.noreply.github.com>
* feat(shared): add enhance-faces tool definition and i18n strings
* feat(ai): add face enhancement script with GFPGAN and CodeFormer support
Detects faces via MediaPipe dual-model approach, then enhances using
GFPGAN (proven) or CodeFormer (via codeformer-pip) with auto fallback.
Supports strength-based alpha blending with original image.
* feat(ai): add TypeScript bridge for face enhancement
* feat(api): add enhance-faces route with GFPGAN/CodeFormer support
* feat(web): add enhance-faces settings component and register in tool registry
* feat(docker): add CodeFormer dependency and model download
- Add codeformer-pip to both CPU and GPU requirements
- Download CodeFormer model (~375MB) at Docker build time
- Add CodeFormer to smoke test verification
* fix(enhance-faces): address code review findings
- Skip alpha blend for CodeFormer (strength already applied via fidelity weight)
- Hide "only enhance main face" checkbox when Best (CodeFormer) is selected
- Fix sensitivity slider labels (swap More/Fewer faces to match actual behavior)
- Register EnhanceFacesControls in pipeline step settings
- Remove model names from user-facing descriptions
* fix(enhance-faces): fix CodeFormer integration and Docker setup
- Add codeformer-pip install to Dockerfile with --no-deps to avoid numpy 2.x conflict
- Re-pin numpy==1.26.4 after codeformer-pip install
- Pin codeformer-pip==0.0.4 in requirements files
- Broaden auto-mode fallback to catch any Exception from CodeFormer
---------
Co-authored-by: stirling-image <stirling-image@users.noreply.github.com>
* feat(noise-removal): register tool in shared constants and i18n
* feat(noise-removal): add SCUNet and NAFNet model architectures
* feat(noise-removal): add Python denoising engine with 4 quality tiers
* feat(noise-removal): add TypeScript bridge for Python sidecar
* feat(noise-removal): add frontend settings with 4-tier selector
* feat(noise-removal): register in tool registry and pipeline
* feat(noise-removal): add Fastify API route with Zod validation
* feat(noise-removal): add SCUNet and NAFNet model downloads to Docker build
* test(noise-removal): add to e2e tool page rendering tests
* test(noise-removal): add integration tests for API endpoint
* style: fix biome formatting and import ordering
* fix(noise-removal): use correct model download URLs
NAFNet model is hosted on HuggingFace, not GitHub releases.
Also align SCUNet URL to use the KAIR releases (same as Docker build).
* fix(noise-removal): remove emojis from tier selector, simplify labels
Drop emoji icons from Quick/Balanced/Quality/Maximum buttons. Replace
technical algorithm names with plain descriptions users can understand.
---------
Co-authored-by: stirling-image <stirling-image@users.noreply.github.com>
