Komponenten

Restore Face (GFPGAN)

@tencentarc

GFPGAN blind face restoration on an Image. Faces are detected, restored and pasted into the upscaled frame; the sink picks the output arm: the restored Image, a UInt64 face count, or the per-face crops.

Inpaint Image (OpenCV)

@opencv

Fills masked holes in a frame with classical OpenCV inpainting: it takes a source Image and a same-size mask Image and emits the repaired Image, choosing Fast Marching or Navier-Stokes. Best for small scratches, logos, or thin watermarks.

Detect Objects (HuggingFace)

@huggingface

Detects objects in each Image with a HuggingFace or Ultralytics checkpoint chosen at load time from the repo contents and projects the detections onto the connected output shape: boxes, geometry, classes, a count, or an overlay. Filtered by confidence.

Extract Text (Surya OCR)

@datalab

Runs Surya detection and recognition over each Image, emitting either a String transcript or structured regions with Polygon or Rectangle bounds, with math recognition and reading-order sorting.

Convert Image Format

Colour-layout converter for Image payloads. Re-encodes each input into the layout set by color_model — BGR, BGRA, GRAY, RGB, or RGBA. Use when downstream expects a specific channel layout different from what upstream emits.

Track Objects (ByteTrack)

@pipelogic

ByteTrack multi-object tracker over per-frame [BoundingBox] or [Segmentation]. Its two-stage association keeps low-confidence detections instead of discarding them, recovering objects through occlusion. A Kalman motion model plus class-aware gating give stable ids. Emits one active-track list per frame, not per box.

Track Objects (BoxMOT)

@boxmot

Detector-agnostic multi-object tracker over the BoxMOT tracker zoo (appearance/ReID or motion-only). Links per-frame [BoundingBox] detections into stable track ids. Track-centric resolved oneof: [Track] (default), tracked objects, boxes, or ids.

Generate Text (HuggingFace)

@huggingface

Generates a reply for each String prompt by running a HuggingFace causal LM in-process with transformers (no vLLM / Ollama / SGLang sidecar). The repo's own chat template formats each turn, with an optional system prompt, and emits the assistant text as a String.

Analyze Image (SGLang VLM)

@lmsys

Sends one or more images plus a String prompt to a vision-language model served by a local SGLang sidecar and returns the model's reply as a String. The image input is a oneof spanning a single image, an image-with-label, a batch, or labeled batch.

Input CSV File

Parses a delimited text file and emits one Analytics.TabularDataset per file, handling quoted cells and an optional header row. Use it to pull a CSV or TSV corpus into a tabular pipeline.

Pack Union

Merges N independently-typed input streams t1, t2, ... into a single union t1|t2, tagging each emitted value with the variant of the input it came from. Inputs are alternatives, not co-arriving fields. Inverse of unpack_union.

Extract Text (TrOCR)

@microsoft

Recognises the text in a single cropped Image line with a TrOCR vision encoder-decoder and emits either a bare String transcript or a {text: String, confidence: Double} record pairing the transcript with the mean per-token recogniser score.

Detect Oriented Objects (Ultralytics YOLO)

@ultralytics

Runs a bound Ultralytics YOLO-OBB checkpoint over each Image with rotated NMS. Projects the per-frame rotated boxes as [OrientedBoundingBox] (default), [Rectangle], [DetectedClass], a UInt64 count, or the annotated Image.

Segment Image (Ultralytics YOLO)

@ultralytics

Real-time instance segmentation on an Image with an Ultralytics YOLO weights file. The sink resolves the projection: mask-plus-box pairs, masks, boxes, parallel arrays, a merged Mask, a UInt64 count, or a colourised label Image.

Detect Faces (Ultralytics YOLO)

@ultralytics

YOLO face detector loaded from a mounted Ultralytics checkpoint (the model weights slot): takes an Image and emits boxes, boxes paired with landmarks, landmark lists, or parallel box/landmark arrays depending on the downstream connection. Use to find faces.

Visualize Expression

Variadic renderer that evaluates a Python expression with all inputs bound to x0, x1, ... (NumPy/OpenCV/Matplotlib in scope) and emits the result as Image. Use for one-off visualisations; for standard overlays prefer visualize.

Track Objects (SAM 3)

@meta

SAM 3 seed-and-propagate video tracker. A detector seeds objects (boxes or masks); SAM 3 memory propagates each object's mask across frames with a stable id. Mask-native resolved oneof: [TrackedSegmentation] (default), tracks, objects, boxes, or ids.

Output Text HTTP

Emits each upstream String (or every element of a [String]) over the channels in transports: a WebSocket broadcast on port 8999, an SSE stream on the same path, and/or a text/plain HTTP POST to url, each channel independent.