Components

Detect Objects (Open Vocabulary Grounding DINO)

@idea_research

Detects objects in each Image from free-text class phrases via a Grounding DINO open-set detector and projects detections onto the connected output: boxes, geometry, classes, a count, or an overlay. Offline from the bound model.

Upscale Image (OpenCV)

@opencv

Enlarges an Image by an integer factor using a classical CPU interpolation kernel (nearest, bilinear, bicubic, or Lanczos) and emits the resized Image; fast and detail-preserving but not generative.

Label COCO Wholebody Hand Landmarks

@bleenco

Reindexes a flat [Landmark] hand list into a named Landmarks2d.Hand21 record (wrist plus four joints per finger: thumb, forefinger, middle, ring, pinky). A Maybe<Landmarks2d.Hand21> sink yields an empty result when the list is short.

Unpack Union

Splits a union-typed stream t1|t2 into one output per variant. Each tick exactly one output fires — the branch matching the message's variant tag — carrying the unwrapped value; the others stay quiet. A router, not a fan-out. Inverse of pack_union.

Input TOML File

Parses TOML files from the bound directory and decodes each document directly into the typed t value the downstream port demands, optionally looping. Use it to ingest TOML configuration into typed pipeline stages.

Lift Unroll

Unrolls N parallel equal-length lists (each [$ts]) into a UInt64 length channel plus N element streams (each $ts), ticking once per element index. Inverse of lift_reroll, which consumes the length to regroup.

Summarize Text (HuggingFace)

@huggingface

Abstractive text summarizer that loads any HuggingFace seq2seq summarization model named by model in-process and condenses a String document into a shorter summary String, bounded by min_length and max_length. No cloud API.

Input Parquet File

Reads Apache Parquet files from the bound directory and emits one Analytics.TabularDataset per file, optionally projecting a subset of columns. Use it to pull a columnar dataset into a tabular pipeline.

Output SRT File (SubRip Text)

Renders each upstream [Subtitle] list to a SubRip .srt file under output_path, formatting cue timings as HH:MM:SS,mmm and naming the file from name_template plus a 6-digit index.

Detect Objects Zero Shot (HuggingFace)

@huggingface

Detects objects named by labels text queries in each Image with an open-vocabulary HuggingFace detector (Grounding DINO / OWLv2) and projects detections onto the connected output shape: boxes, geometry, classes, a count, or an overlay.

Generate Speech (Chatterbox)

@resembleai

Turns each String utterance into a mono AudioFrame using the ResembleAI Chatterbox-Turbo single-step decoder, honouring inline paralinguistic tags (cough, laugh, chuckle) written directly in the text.

Extract Image Embedding (OpenCLIP)

@laion

Passes each Image through an OpenCLIP image tower — the architecture arch built to match the weights bound to model — and emits one Embedding, optionally L2-normalised for cosine or inner-product retrieval. Use for image-side retrieval, clustering, or similarity; pair with a matching CLIP text encoder for cross-modal search.

Upscale Audio (AudioSR)

@haoheliu

Reconstructs the missing high band of a 2-16 kHz AudioFrame with the AudioSR latent-diffusion model and emits a 48 kHz mono AudioFrame; output rate and layout are fixed regardless of the input.

Label Animal Landmarks (AP 10K)

@bleenco

Maps a flat AP-10K-ordered [Landmark] list onto a typed Landmarks2d.Animal17 record with named fields — eyes, nose, neck, tail, shoulders, elbows, paws, hips, and knees — so downstream stages address keypoints by name instead of index.

Input Video File

Decodes video files from the bound directory and emits a VideoFrame stream at each source's native frame rate, optionally resizing and looping. Use it for offline replay or deterministic tests against recorded footage.

Generate Speech (Kokoro)

@hexgrad

Synthesises each String into a fixed 24 kHz mono AudioFrame with the Kokoro local TTS model, concatenating every segment Kokoro emits for the input into one continuous waveform.

Select Stream

Routes one of N same-typed candidate streams t to the output, chosen each tick by a leading UInt64 index. The index controls routing only; the output is the selected candidate of type t, forwarded unchanged.

Detect Landmarks (HuggingFace)

@huggingface

Crops each box of a [BoundingBox] list out of an Image and runs a HuggingFace pose/keypoint checkpoint (ViTPose, SuperPoint, or ONNX) to emit per-detection landmarks as [[Landmark]].