Components

Pack Record

Packs N positional input streams, each carrying a per-field type $ts, into a single record {$ns: $ts...}, assigning each input to a named field in declared order with its type preserved. Inverse of unpack_record.

Normalize Audio Loudness (EBU R128)

@jiixyj

Measures the integrated loudness of an AudioFrame, scales it to a target LUFS, and limits the resulting true peak, emitting a leveled AudioFrame with unchanged sample rate and channel count for consistent downstream levels.

Detect Objects (TorchServe)

@pytorch

JPEG-encodes each Image and posts it to a TorchServe-hosted detection model, parsing the JSON response into one per-detection record, then projecting that record onto the connected output shape: boxes, geometry, labels, a count, or an overlay, filtered by confidence.

Restore Document (DocDiff)

@royalvice

Residual-diffusion document restoration on a degraded page Image. A coarse pass plus a short residual diffusion sharpen blurred or noised glyphs while keeping layout intact, emitting a restored Image at the input resolution. Failures pass the input through.

Crop Image

Crops an Image to a region given as a BoundingBox, Rectangle, Polygon, Mask, or Ellipse, returning either a tight crop or the original-size frame with outside pixels painted by a configurable fill colour.

Unpack Record

Splits a record-typed stream into one output per field, in declared field order. Field names are dropped; each output port carries that field's type $ts verbatim. Arity equals the number of declared fields. Inverse of pack_record.

Forecast Time Series (Chronos)

@amazon

Zero-shot forecaster on a bound pretrained Chronos checkpoint. Reads the numeric target_column of an Analytics.ForecastInput, feeds trailing history, and emits a Analytics.ForecastResult (or point list) with Chronos-derived intervals.

Detect Document Layout (DocLayout YOLO)

@opendatalab

Detects layout regions (title, text, table, figure, formula, caption) on a page Image with DocLayout-YOLO and emits one detection per region as [BoundingBox] (with class id) or geometry-only [Rectangle], fixed at deploy by the consumer.

Diarize Speakers (pyannote)

@pyannote

Runs the pyannote community-1 speaker-diarization pipeline bound at the model slot over each AudioFrame, emitting per-turn [{start: Double, end: Double, speaker: String}] records — who spoke when — upstream of transcription to split audio into per-speaker turns.

Unpack Named

Strips a named-type stream named_type down to its underlying structural shape unwrapped_named_type — the alias's declared body. Type-level relabel only; the value passes through unchanged. Inverse of pack_named.

Classify Image (TorchServe)

@pytorch

Classifies each Image through a TorchServe-hosted model archive; the ranked result projects onto the connected arm: Classification, [Classification], String, [DetectedClass], UInt64, or [Double].

Upscale Image (Swin2sr)

@mvlab

Super-resolves an Image with the Swin2SR transformer at the factor inferred from the chosen checkpoint and emits the enlarged Image; the default variant targets real-world noisy or compressed photos.

Track Bounding Boxes (Hungarian Algorithm)

@bleenco

Hungarian-algorithm multi-object tracker over per-frame [BoundingBox] or [Segmentation]. Class-aware gating, a tentative/confirmed/coasting lifecycle, and motion prediction give stable ids through occlusions. Emits one active-track list per frame, not per box.

Track Objects (Ultralytics YOLO)

@ultralytics

Combined Ultralytics YOLO detector and tracker. Accepts an Image per tick, runs detection with persistent tracker memory, and emits [(BoundingBox, UInt64)] stable-id pairs for all matched objects.

Estimate Metric Depth (Depth Anything V3)

@bytedance

Metric monocular depth from Depth Anything 3: routes each Image through a DA3 metric head and emits a deploy-time-resolved greyscale depth Image or a MetricDepthMap in metres, recovering scale from a predicted or assumed focal length.

Transcribe Audio (Moonshine)

@moonshine

UsefulSensors Moonshine ASR. Resamples each AudioFrame to 16 kHz and decodes it to a String. Each checkpoint is monolingual, so the language is chosen by the model you load; for in-stream multilingual audio prefer transcribe_audio_faster_whisper.

Input VTT File

Parses WebVTT subtitle files from the bound directory and emits the whole file as one [Subtitle] list per document, matching the output of input_srt_file_subrip_text. Use it when the captions arrive in WebVTT.

Classify Document (Donut)

@naver

Reads a whole page Image end-to-end with an OCR-free Donut encoder-decoder and decodes the document type directly, emitting a single Classification whose confidence is the mean per-token probability of the generated label.