Komponenten

Extract Text (PaddleOCR VL)

@paddlepaddle

PaddleOCR-VL document parsing. Each Image is parsed on the selected device with layout detection and vision-language recognition, emitting a joined String transcript or per-block regions with boxes. SOTA parser (PaddleOCR-VL-1.6, 0.9B); CUDA GPU required.

Unwarp Document (UVDoc)

@tanguymagne

Rectifies a curved, folded, or perspective-distorted document Image with the UVDoc grid network, which regresses a 2D unwarping grid in a single forward pass, then remaps the original frame through it and emits the flattened Image for OCR.

Serialize to JSON

Encodes any typed value t as a one-line JSON String. Use it between a typed producer and a textual sink — an HTTP, messaging or log target that expects a string payload rather than a typed value.

Merge Streams

Fans several same-typed input streams t into one output of the same type, interleaving each value as it fires. Values pass through unchanged; this is a merge via merge_streams, not a join — inputs are not paired or zipped.

Analyze Document Layout (Donut)

@naver

OCR-free document understanding with a Donut model: reads one document Image end-to-end and emits a downstream-resolved oneof t[String, [(String, String)]] — either the raw decoded sequence as text, or the structured parse flattened to dotted key/value pairs.

Summarize Tracked Object

@pipelogic

Per-track Bayesian classifier stabiliser. Reads per-frame class [Double] probabilities tagged with a track id, blends them over a sliding history weighted by confidence and prior, and emits oneof t[[(UInt64, UInt64, Bool)], [Observation.Event]].

Detect Landmarks (Triton)

@nvidia

Crops each box of a [BoundingBox] list from an Image and runs a Triton-served pose checkpoint per box, decoding heatmaps or SimCC into per-detection keypoints emitted as [[Landmark]].

Track Objects (BoT SORT)

@pipelogic

BoT-SORT multi-object tracker over per-frame [BoundingBox] or [Segmentation]. A ByteTrack-style two-stage IoU association keeps low-confidence detections instead of discarding them, a Kalman motion model predicts each track, and a per-frame camera-motion estimate warps prior boxes so global camera movement does not break the match. Emits one active-track list per frame, not per box.

Animate Portrait Audio (SadTalker)

@tencent

Audio-driven talking head with SadTalker: consumes a driving AudioFrame speech stream, extracts 3D head-pose and expression motion, and renders a bound source portrait into a sequence of Image frames that form the talking-head video.

Classify Text (Zero Shot)

@moritzlaurer

Scores a String against a runtime-supplied set of candidate labels with an offline zero-shot NLI model, emitting the ranked labels as [Classification], the single best as Classification, or the raw per-label [Double] scores.

Output Video File

Encodes each upstream VideoFrame into a rolling set of video files on disk, starting a fresh segment every file_size_seconds and keeping at most max_files. The container and encoder follow format and video_codec.

Output CSV File

Writes each Analytics.TabularDataset to a fresh character-separated file named from a template plus a zero-padded index, with a configurable single-char separator, persisting a stream of typed tables to disk.

Embed Face (AdaFace)

@minchul

Takes an Image plus 5-point [[Landmark]] sets, similarity-aligns each face to a canonical crop, and runs an AdaFace recognizer to emit one Embedding per face — quality-adaptive margins improve accuracy on low-quality or occluded faces versus ArcFace.

Denoise Audio (ClearerVoice Studio)

@modelscope

ClearVoice speech enhancement. Accepts AudioFrame (flattened to mono), resamples to model_name native rate (16 or 48 kHz), and emits a denoised AudioFrame. Use upstream of ASR, VAD, or diarisation when quality matters more than latency.

Estimate Depth (HuggingFace)

@huggingface

Monocular depth estimator over each Image that auto-selects a HuggingFace pipeline, eager transformers, or ONNX backend from the checkpoint and emits a deploy-time-resolved depth Image, a metric MetricDepthMap, or an unscaled relative depth map.

Segment Cells (Cellpose SAM)

@mouseland

Generalist cell instance segmentation on a microscopy Image with Cellpose-SAM. The sink resolves the projection: mask-plus-box pairs, masks, boxes, parallel arrays, a merged Mask, a UInt64 count, or a colourised label Image.

Classify Speech Emotion (Zero Shot)

@laion

Embeds each AudioFrame and a runtime-supplied set of emotion labels into one CLAP space and ranks the labels by audio-text similarity, emitting [Classification], the raw [Double] scores, or just the best label as a String.

Visualize Vector Field

Optical-flow renderer. Converts a dense OpticalFlow field into a Middlebury color-wheel Image (color = direction, intensity = magnitude). Magnitude is normalized per frame; use for inspecting or debugging dense optical-flow outputs.