AI Engine Reference

The @snapotter/ai package coordinates native tools and Python runtimes for local ML operations. Most ML tools use a persistent Python sidecar for fast warm starts. OCR is intentionally separate: fast invokes the native Tesseract binary, while balanced and best use a dedicated persistent JSONL dispatcher pinned to the active immutable RapidOCR generation under /data/ai/v3. Each request holds a generation lease. During an upgrade, SnapOtter runs a smoke test on the candidate before activation, atomically switches to the new dispatcher, then drains the old generation before garbage collection.

NVIDIA CUDA is auto-detected and used by runtimes that support it. OCR uses CPU on every host, including systems with NVIDIA GPUs, avoiding CUDA and driver coupling for this tool.

Intel/AMD iGPU acceleration through VA-API, Quick Sync, or OpenCL is not supported for AI inference today. Mapping /dev/dri into a container does not accelerate these Python sidecar tools unless a CUDA-capable NVIDIA GPU is available.

19 Python sidecar AI tools across four modalities (image, audio, video, document), plus 2 tools with optional AI capabilities. All models run locally - no internet required after initial model download.

Architecture

Node.js Tool Route
      |
      v
 @snapotter/ai bridge.ts
      | (stdin/stdout JSON + stderr progress events)
      v
 +-- Native Tesseract + Ghostscript (fast image/PDF OCR)
 |
 +-- Isolated OCR runtime (persistent JSONL dispatcher)
 |     `-- RapidOCR + ONNX Runtime CPU + pinned PP-OCR models
 |
 `-- Python dispatcher (persistent process, "ai" profile)
      |
      |-- remove_bg.py        (rembg / BiRefNet)
      |-- upscale.py          (RealESRGAN)
      |-- inpaint.py          (LaMa ONNX)
      |-- outpaint.py         (LaMa canvas expansion)
      |-- detect_faces.py     (MediaPipe)
      |-- face_landmarks.py   (MediaPipe landmarks)
      |-- enhance_faces.py    (GFPGAN / CodeFormer)
      |-- colorize.py         (DDColor)
      |-- noise_removal.py    (SCUNet / tiered denoising)
      |-- red_eye_removal.py  (landmark + color analysis)
      |-- restore.py          (scratch repair + enhancement + denoising)
      |-- transcribe.py       (faster-whisper speech-to-text)
      +-- install_feature.py  (on-demand bundle installer)

A separate "docs" dispatcher profile replaces the AI allowlist with document-processing scripts (doc_pagecount, doc_health, doc_flatten, doc_redact, doc_text, doc_to_word, doc_metadata, doc_html_pdf) and skips heavy ML imports.

Timeouts: 300 s default; OCR and BiRefNet background removal get 600 s.

Feature Bundles

AI models are packaged by shared dependency stack, not one archive per tool. A feature bundle can enable several tools when they use the same model family, Python wheels, or native libraries. This keeps the release Docker image smaller and avoids storing duplicate copies of the same background matting, face detection, OCR, restoration, and speech models.

The Docker image ships the application plus the common runtime. Large model archives are downloaded on demand into the persistent /data/ai volume, then reused by every tool that needs them. If a bundle is already installed because another tool needed it, enabling a new dependent tool does not download that bundle again.

Most AI tools require one or more feature bundles before they can run. The admin UI installs those by tool through POST /api/v1/admin/tools/:toolId/features/install, which resolves the full bundle list, skips bundles that are already installed, and queues only the missing downloads. For example, enabling Passport Photo on a fresh instance queues background-removal and face-detection; enabling it after Background Removal is already installed queues only face-detection. OCR is the exception because fast needs no pack; install its optional accurate runtime through the UI or POST /api/v1/admin/features/ocr/install.

Bundle	Size	Shared dependency group	Tools that use it
background-removal	4-5 GB	rembg / BiRefNet background matting	remove-background, passport-photo, transparency-fixer, background-replace, blur-background
face-detection	200-300 MB	MediaPipe face detection and landmarks	blur-faces, red-eye-removal, smart-crop
object-eraser-colorize	1-2 GB	LaMa inpainting/outpainting and DDColor	erase-object, colorize, ai-canvas-expand
upscale-enhance	5-6 GB	RealESRGAN, GFPGAN / CodeFormer, denoising	upscale, enhance-faces, noise-removal
photo-restoration	4-5 GB	scratch repair and restoration pipeline	restore-photo
ocr	~208-234 MiB download / ~409-488 MiB installed	Optional RapidOCR 3.9.1, ONNX Runtime 1.20.1, and pinned PP-OCR models	ocr, ocr-pdf (balanced and best only)
transcription	~600 MB	faster-whisper speech-to-text models	transcribe-audio, auto-subtitles

Tools with cross-bundle dependencies:

Tool	Required bundles	Why
passport-photo	background-removal, face-detection	Removes the background, then uses face landmarks to frame the crop to passport and ID photo rules.
enhance-faces	upscale-enhance, face-detection	Detects faces before running GFPGAN or CodeFormer enhancement on the selected face regions.

A tool is available only when all of its required bundles are installed, except OCR: its built-in fast tier remains available without the optional OCR pack. Partial installs are valid and are handled incrementally: installed bundles are reused, missing bundles are shown as downloads, and queued installs run one at a time so the shared Python environment is not modified concurrently.

Accurate OCR runtime installation

The accurate OCR pack is a platform-specific runtime for the official Linux amd64 or Linux arm64 container. The amd64 build uses Python 3.12; the arm64 build uses Python 3.11. Both builds run RapidOCR through ONNX Runtime's CPUExecutionProvider, so the same pack works on CPU-only and NVIDIA Docker hosts. The accurate runtime requires at least 4 GiB of effective memory: the configured container cgroup limit, otherwise host memory. A system below that signed compatibility minimum is rejected before download. This requirement does not apply to built-in Fast OCR. Bare-metal builds are rejected because their libc and Python ABI cannot be inferred safely; Fast OCR remains available when the host provides Tesseract and Ghostscript. Fast supports auto, en, de, es, fr, zh, and ja, but not Korean (ko). Korean therefore requires a supported accurate runtime and a balanced or best tier; unsupported hosts receive an explicit incompatibility response rather than a silent Fast fallback.

The optional artifact is about 208-234 MiB compressed and 409-488 MiB extracted, depending on architecture. The signed index binds the exact compressed and extracted byte counts enforced by the installer. Built-in Tesseract adds about 25 MiB to the official image and needs no files in /data/ai.

Online installation fetches a signed release index and the exact content-addressed artifact for the current platform. SnapOtter verifies the Ed25519 index signature, artifact size, SHA-256 digest, model digests, paths, file modes, and staged smoke test before atomically activating the new generation. A failed install leaves the prior healthy generation active.

For air-gapped installation, upload both the release's ocr-runtime-index.json and matching OCR runtime archive to POST /api/v1/admin/features/import using multipart fields named index and archive. Offline import applies the same signature, hash, extraction, compatibility, and smoke-test checks as online installation; an archive without its trusted signed index is rejected.

---

Background Removal

Tool route: remove-backgroundModel: rembg with BiRefNet (default) or U2-Net variants

Parameter	Type	Default	Description
model	string	-	Model variant (optional override)
backgroundType	string	"transparent"	One of: transparent, color, gradient, blur, image
backgroundColor	string	-	Hex color for solid background
gradientColor1	string	-	First gradient color
gradientColor2	string	-	Second gradient color
gradientAngle	number	-	Gradient angle in degrees
blurEnabled	boolean	-	Enable background blur effect
blurIntensity	number (0-100)	-	Blur intensity
shadowEnabled	boolean	-	Enable drop shadow on subject
shadowOpacity	number (0-100)	-	Shadow opacity
outputFormat	string	-	Output format: png, webp, or avif
edgeRefine	integer (0-3)	-	Edge refinement level
decontaminate	boolean	-	Remove color bleed from edges

Background Replace

Tool route: background-replaceModel: rembg / BiRefNet (shared with remove-background)

Removes the background and replaces it with a solid color or gradient.

Parameter	Type	Default	Description
backgroundType	"color" | "gradient"	"color"	Background mode
color	string	"#ffffff"	Background hex color (when backgroundType is color)
gradientColor1	string	-	First gradient hex color
gradientColor2	string	-	Second gradient hex color
gradientAngle	integer (0-360)	180	Gradient angle in degrees
feather	integer (0-20)	0	Edge feathering radius
format	"png" | "webp"	"png"	Output format

Blur Background

Tool route: blur-backgroundModel: rembg / BiRefNet (shared with remove-background)

Blurs the background while keeping the subject sharp.

Parameter	Type	Default	Description
intensity	integer (1-100)	50	Blur intensity
feather	integer (0-20)	0	Edge feathering radius
format	"png" | "webp"	"png"	Output format

Image Upscaling

Tool route: upscaleModel: RealESRGAN (with Lanczos fallback when unavailable)

Parameter	Type	Default	Description
scale	number	2	Upscale factor
model	string	"auto"	Model variant
faceEnhance	boolean	false	Apply GFPGAN face enhancement pass
denoise	number	0	Denoising strength
format	string	"auto"	Output format override
quality	number	95	Output quality (1-100)

OCR / Text Extraction

Tool route: ocrModels: Tesseract (fast); RapidOCR with PP-OCRv6 small models (balanced); PP-OCRv6 medium models with calibrated variant scoring (best)

Parameter	Type	Default	Description
quality	"fast" | "balanced" | "best"	Dynamic	Processing tier. Omitted quality selects the highest available tier in this order: best, balanced, fast. Korean never selects fast; without an accurate tier it returns the accurate-runtime install or compatibility error
language	string	"auto"	Language: auto, en, de, fr, es, zh, ja, ko. Fast does not support ko
enhance	boolean	Tier-dependent	Improve local contrast. Fast applies it directly; accurate tiers keep the variant only when calibrated scoring improves OCR. Defaults on for Best
engine	string	-	Deprecated compatibility alias. Maps tesseract to fast and the legacy paddleocr value to balanced; it does not load PaddlePaddle

Returns extracted text plus provenance metadata: engine, requested and actual quality, device, provider, degradation state, warnings, and accurate-runtime/model versions when applicable. Explicit quality requests never fall back to another tier. If balanced or best is unavailable, the API returns FEATURE_NOT_INSTALLED or FEATURE_INCOMPATIBLE instead of silently running fast. Explicit Fast or legacy tesseract with Korean returns FEATURE_INCOMPATIBLE, compatibilityReason: "fast-korean-unsupported", and accurate-pack guidance before a job is queued.

PDF OCR

Tool route: ocr-pdfModels: Same tier system as image OCR

Extracts text from scanned PDF documents using AI-powered OCR, page by page.

Parameter	Type	Default	Description
quality	"fast" | "balanced" | "best"	Dynamic	Processing tier. Omitted quality selects the highest available tier in this order: best, balanced, fast. Korean never selects fast; without an accurate tier it returns the accurate-runtime install or compatibility error
language	string	"auto"	Language: auto, en, de, fr, es, zh, ja, ko. Fast does not support ko
pages	string	"all"	Page selection: "all", "1-3", "1,3,5"
enhance	boolean	Tier-dependent	Improve local contrast. Fast applies it directly; accurate tiers keep the variant only when calibrated scoring improves OCR. Defaults on for Best
engine	string	-	Deprecated compatibility alias. Maps tesseract to fast and the legacy paddleocr value to balanced; it does not load PaddlePaddle

The same no-downgrade and Korean compatibility rules apply to PDF OCR. PDF pages are rasterized before recognition, and one request can select at most 50 pages.

Face / PII Blur

Tool route: blur-facesModel: MediaPipe face detection

Parameter	Type	Default	Description
blurRadius	number (1-100)	30	Gaussian blur radius
sensitivity	number (0-1)	0.5	Detection confidence threshold

Face Enhancement

Tool route: enhance-facesModels: GFPGAN, CodeFormer

Parameter	Type	Default	Description
model	"auto" | "gfpgan" | "codeformer"	"auto"	Enhancement model
strength	number (0-1)	0.8	Enhancement strength
sensitivity	number (0-1)	0.5	Face detection threshold
onlyCenterFace	boolean	false	Enhance only the most central face

AI Colorization

Tool route: colorizeModel: DDColor (with OpenCV DNN fallback)

Converts black-and-white or grayscale photos to full color.

Parameter	Type	Default	Description
intensity	number (0-1)	1.0	Color saturation strength
model	"auto" | "ddcolor" | "opencv"	"auto"	Model variant

Noise Removal

Tool route: noise-removalModel: SCUNet (tiered denoising pipeline)

Parameter	Type	Default	Description
tier	"quick" | "balanced" | "quality" | "maximum"	"balanced"	Processing tier
strength	number (0-100)	50	Denoising strength
detailPreservation	number (0-100)	50	How much detail to preserve; higher keeps more texture
colorNoise	number (0-100)	30	Color noise reduction strength
format	string	"original"	Output format: original, png, jpeg, webp, avif, jxl
quality	number (1-100)	90	Output encoding quality

Red Eye Removal

Tool route: red-eye-removal

Detects face landmarks, locates eye regions, and corrects red-channel oversaturation.

Parameter	Type	Default	Description
sensitivity	number (0-100)	50	Red pixel detection threshold
strength	number (0-100)	70	Correction strength
format	string	-	Output format override (optional)
quality	number (1-100)	90	Output quality

Photo Restoration

Tool route: restore-photo

Multi-step pipeline for old or damaged photos: scratch/tear detection and repair, face enhancement, denoising, and optional colorization.

Parameter	Type	Default	Description
scratchRemoval	boolean	true	Detect and repair scratches, tears
faceEnhancement	boolean	true	Apply face enhancement pass
fidelity	number (0-1)	0.7	Face enhancement strength (higher = more conservative)
denoise	boolean	true	Apply denoising pass
denoiseStrength	number (0-100)	25	Denoising strength
colorize	boolean	false	Colorize after restoration
colorizeStrength	number (0-100)	85	Colorization intensity

Passport Photo

Tool route: passport-photoModels: MediaPipe face landmarks + BiRefNet background removal

Two-phase workflow: analyze (detect face + remove background) then generate (crop, resize, tile). Supports 37+ countries across 6 regions.

Phase 1: Analyze

POST /api/v1/tools/image/passport-photo/analyze

Accepts an image file (multipart). Returns face landmark data, a base64 preview, and image dimensions.

Phase 2: Generate

POST /api/v1/tools/image/passport-photo/generate

Accepts a JSON body with the Phase 1 results plus generation settings:

Parameter	Type	Default	Description
jobId	string	(required)	Job ID from Phase 1
filename	string	(required)	Original filename from Phase 1
countryCode	string	(required)	ISO country code (e.g., US, GB, IN)
documentType	string	"passport"	Document type
bgColor	string	"#FFFFFF"	Background color hex
printLayout	string	"none"	Print layout: none, 4x6, a4, letter
maxFileSizeKb	number	0	Max file size in KB (0 = no limit)
dpi	number (72-1200)	300	Output DPI
customWidthMm	number	-	Custom width in mm (overrides country spec)
customHeightMm	number	-	Custom height in mm (overrides country spec)
zoom	number (0.5-3)	1	Zoom factor
adjustX	number	0	Horizontal position adjustment
adjustY	number	0	Vertical position adjustment
landmarks	object	(required)	Landmarks from Phase 1
imageWidth	number	(required)	Image width from Phase 1
imageHeight	number	(required)	Image height from Phase 1

Object Erasing (Inpainting)

Tool route: erase-objectModel: LaMa via ONNX Runtime

The mask is sent as a second file part (fieldname mask), not as base64. White pixels in the mask indicate areas to erase. The format and quality settings are sent as top-level form fields.

Parameter	Type	Default	Description
file	file	(required)	Source image (multipart)
mask	file	(required)	Mask image (multipart, fieldname mask, white = erase)
format	string	"auto"	Output format: auto, png, jpg, jpeg, webp, tiff, gif, avif, heic, heif, jxl
quality	integer (1-100)	95	Output quality

CUDA-accelerated when an NVIDIA GPU is available.

AI Canvas Expand

Tool route: ai-canvas-expandModel: LaMa-based outpainting

Expands the canvas of an image in any direction and fills new areas with AI-generated content that matches the existing image.

Parameter	Type	Default	Description
extendTop	integer	0	Pixels to extend at the top
extendRight	integer	0	Pixels to extend at the right
extendBottom	integer	0	Pixels to extend at the bottom
extendLeft	integer	0	Pixels to extend at the left
tier	"fast" | "balanced" | "high"	"balanced"	Quality tier
format	string	"auto"	Output format: auto, png, jpg, jpeg, webp, tiff, gif, avif, heic, heif, jxl
quality	integer (1-100)	95	Output quality

At least one extend direction must be greater than 0.

Smart Crop

Tool route: smart-cropModel: MediaPipe face detection (face mode only)

Parameter	Type	Default	Description
mode	string	"subject"	Crop strategy: subject, face, trim
strategy	"attention" | "entropy"	"attention"	Strategy for subject mode
width	integer	-	Output width
height	integer	-	Output height
padding	integer (0-50)	0	Padding percentage around subject
facePreset	string	"head-shoulders"	Preset framing when mode=face
sensitivity	number (0-1)	0.5	Face detection threshold
threshold	integer (0-255)	30	Background detection threshold (trim mode)
padToSquare	boolean	false	Pad trimmed result to a square
padColor	string	"#ffffff"	Background color for square padding
targetSize	integer	-	Target size for padded output (pixels)
quality	integer (1-100)	-	Output quality

Legacy mode values attention and content are accepted and mapped to subject and trim respectively.

Face presets:

Preset	Best for
closeup	Headshots
head-shoulders	Profile photos
upper-body	LinkedIn / formal
half-body	Full upper body

Transcribe Audio

Tool route: transcribe-audioModel: faster-whisper

Converts speech to text. Supports plain text, SRT, and VTT output formats.

Parameter	Type	Default	Description
language	string	"auto"	Language: auto, en, de, fr, es, zh, ja, ko, id, th, vi
outputFormat	"txt" | "srt" | "vtt"	"txt"	Output format

Auto Subtitles

Tool route: auto-subtitlesModel: faster-whisper (extracts audio from video, then transcribes)

Generates subtitle files from a video's audio track.

Parameter	Type	Default	Description
language	string	"auto"	Language: auto, en, de, fr, es, zh, ja, ko, id, th, vi
format	"srt" | "vtt"	"srt"	Output subtitle format

PNG Transparency Fixer

Tool route: transparency-fixerModel: BiRefNet HR-matting (2048x2048 resolution)

Fixes "fake transparent" PNGs where the background was removed but left behind fringing, halos, or semi-transparent artifacts. Uses BiRefNet's high-resolution matting model to produce a clean alpha channel, then applies configurable defringe processing to remove color contamination along edges.

OOM fallback chain: If BiRefNet HR-matting exceeds available memory, the tool automatically falls back to birefnet-general, then to u2net.

Parameter	Type	Default	Description
defringe	number (0-100)	30	Edge defringe strength to remove color contamination
outputFormat	"png" | "webp"	"png"	Output image format
removeWatermark	boolean	false	Apply watermark removal pre-processing (median filter)

curl -X POST http://localhost:1349/api/v1/tools/image/transparency-fixer \
  -H "Authorization: Bearer <token>" \
  -F "[email protected]" \
  -F 'settings={"defringe":30,"outputFormat":"png"}'

---

Tools with Optional AI Capabilities

The following tools are not Python sidecar tools but use AI features when certain options are enabled.

Image Enhancement

Tool route: image-enhancementEngine: Analysis-based (Sharp histogram and statistics)

Analyzes the image and applies automatic corrections for exposure, contrast, white balance, saturation, sharpness, and noise. Supports scene-specific modes.

Parameter	Type	Default	Description
mode	"auto" | "portrait" | "landscape" | "low-light" | "food" | "document"	"auto"	Scene mode for tuning corrections
intensity	number (0-100)	50	Overall correction strength
corrections.exposure	boolean	true	Apply exposure correction
corrections.contrast	boolean	true	Apply contrast correction
corrections.whiteBalance	boolean	true	Apply white balance correction
corrections.saturation	boolean	true	Apply saturation correction
corrections.sharpness	boolean	true	Apply sharpness correction
corrections.denoise	boolean	true	Apply denoising
deepEnhance	boolean	false	Enable AI noise removal via SCUNet (requires upscale-enhance bundle)

An additional analysis endpoint is available at POST /api/v1/tools/image/image-enhancement/analyze which returns the detected corrections without applying them.

Content-Aware Resize (Seam Carving)

Tool route: content-aware-resizeEngine: Go caire binary (not Python - no GPU benefit)

Intelligently resizes images by removing low-energy seams, preserving important content.

Parameter	Type	Default	Description
width	number	-	Target width
height	number	-	Target height
protectFaces	boolean	false	Protect detected face regions (requires face-detection bundle)
blurRadius	number (0-20)	4	Pre-blur for energy calculation
sobelThreshold	number (1-20)	2	Edge sensitivity threshold
square	boolean	false	Force square output