Komponenten

Output NumPy File

Writes each upstream Tensor to a .npy file under output_path, subsampling with every_n and capping retention with max_files. The .npy header carries the tensor shape unless flatten collapses it to 1-D.

Parse JSON

Parses each String as RFC 8259 JSON and emits the typed value t bound at link time, decoding the text into whatever pipelogic type the downstream connection requires. Inverse of serialize_to_json.

Denoise Audio (DeepFilterNet3)

@hendrikschroter

DeepFilterNet 3 full-band speech denoiser: each AudioFrame is mixed to mono, resampled to 48 kHz, cleaned, and emitted as a denoised 48 kHz AudioFrame. Real-time-capable on CPU. Use to clean a noisy speech stream ahead of ASR or VAD.

Extract Text (docTR OCR)

@mindee

Two-stage docTR OCR over each Image; emits a joined String transcript or per-line regions with Polygon or Rectangle bounds (optional confidence). Detection + recognition weights are mounted file links, loaded locally.

Track Objects (OC SORT)

@pipelogic

OC-SORT multi-object tracker over per-frame [BoundingBox] or [Segmentation]. Observation-centric: when a track is re-observed after occlusion its Kalman filter is re-updated along the trajectory between observations, and observation momentum keeps headings consistent, so ids survive the gap. Motion-only, no appearance model. Emits one active-track list per frame, not per box.

Detect Objects (Florence 2)

@microsoft

Runs Florence-2 closed-set detection over each Image, emitting the shape the connected output picks: [BoundingBox], [Rectangle], [DetectedClass], a UInt64 count, or an Image overlay. Offline from the bound model.

Input Mesh File

Imports 3D mesh files from the bound directory with Assimp and emits one Mesh per file at a paced rate, triangulating faces and flattening multi-mesh scenes. Use it for offline batches or 3D asset tests.

Project Points to Plane

@bleenco

Maps [Vector2d] image-space points onto a calibrated ground plane, emitting [Vector2d] world-coordinate positions. Uses a 9-element homography with radial lens undistortion.

Pack Named

Wraps a structurally-typed stream unwrapped_named_type into its named-type identity named_type, producing a strongly-typed alias from its underlying representation. Type-level relabel only; inverse of unpack_named.

Input PDF HTTP

Exposes a public PDF ingress that accepts client-pushed bytes over HTTP POST or WebSocket, parses each upload with Poppler, and emits a oneof t[String, Image, Document, Document.Page] chosen by what it is wired to: a text consumer receives one String per upload, an image consumer receives one Image per page, a Document consumer receives one fully-structured document, a Document.Page consumer receives one structured page streamed per page. Use it when a deployment needs to take PDFs from outside the graph.

Rerank Text (HuggingFace)

@huggingface

Cross-encoder reranking of retrieved passages with any HuggingFace reranker repo loaded in-process. Scores every query-document pair in a RerankInput via AutoModelForSequenceClassification, sorts by descending relevance, optionally truncates to top_n, and emits [RerankResult] carrying each document, score and original index.

Denoise Image (OpenCV)

@opencv

Classical CPU image denoiser over an Image: applies non-local-means, bilateral, Gaussian, or median filtering and returns a cleaned Image in the same color space. No model, no GPU. Pick this for sensor noise, JPEG quantisation, or salt-and-pepper impulse.

Estimate Optical Flow (Triton)

@nvidia

Dense optical flow from any 2-frame model served by NVIDIA Triton (RAFT, GMFlow, FlowFormer, SEA-RAFT and similar exports): takes Image frames and emits per-pixel OpticalFlow against a buffered reference frame, resized to the source dimensions.

Transcribe Audio (HuggingFace)

@huggingface

HuggingFace transformers Whisper transcriber. Resamples each AudioFrame to 16 kHz and decodes it via an ASR pipeline, emitting a String or timed [Subtitle] list. For production speed prefer transcribe_audio_faster_whisper.

Analyze Image (HuggingFace)

@huggingface

Answers a natural-language prompt about each Image using a HuggingFace image-text-to-text model loaded directly in-process with transformers (default Qwen2-VL-2B-Instruct, no vLLM/SGLang/Ollama sidecar) and emits the answer as a String.

Detect Text Regions (CRAFT)

@naver

Runs a CRAFT text-detection checkpoint over each Image and emits one Polygon per detected text region — raw region geometry for cropping or masking rather than extracted text strings.

Separate Speakers (MossFormer2)

@modelscope

Speech separation that splits overlapping voices in an AudioFrame into a [AudioFrame] list with one 16 kHz mono frame per separated speaker. Use it ahead of transcription or capture so each voice is processed on its own clean stream.

Unwarp Document (DocRes)

@jiaxinzhang

Restores a document photo Image with the unified DocRes network and emits a restored Image; one configured objective selects geometric dewarping, deshadowing, illumination cleanup, deblurring, or binarization.