Components

Classify Image (HuggingFace)

@huggingface

Classifies an Image with a HuggingFace, timm, or Optimum network and projects one internal ranked result into the connected output arm: top-1 Classification, ranked [Classification], a String label, and more.

Estimate Depth (Marigold)

@prs_eth

Diffusion-based monocular depth estimator (Marigold) over each Image, emitting a deploy-time-resolved choice of an affine-invariant relative depth map or a greyscale depth Image. The depth is scale- and shift-invariant relative depth, never metric.

Upscale Audio (UniverSR)

@woongzip1

Lifts a band-limited AudioFrame to 48 kHz mono with the vocoder-free UniverSR model, which super-resolves directly in the complex STFT domain via flow matching and reconstructs through an inverse STFT; output rate and layout are fixed.

Input YAML File

Parses YAML files from the bound directory and decodes each document into the typed t value the downstream port demands, streaming multi-document files one record at a time. For TOML use input_toml_file; for JSON Lines use input_jsonl_file.

Extract Formula (PaddleX)

@paddlepaddle

Reads a cropped mathematical formula Image and emits its LaTeX source as a String via the PP-FormulaNet recognition head. Blank or unreadable crops return an empty String. One formula per frame, not a full page.

Output Text File

Writes each upstream String (or every element of a [String]) to disk under output_path. Rotate writes one file per input; append accumulates inputs into a single file joined by line_separator.

Split Text (wtpsplit)

@bminixhofer

Splits a String into a [String] of sentences with a learned segment-any-text (SaT) model. Robust to missing or unusual punctuation and language-agnostic across major scripts, it prepares text for per-sentence downstream processing such as retrieval or generation.

Filter Pose Bounding Box

Estimates yaw from a Landmarks2d.Human22 pose using shoulder spread and spine height and emits Bool true when the subject is frontal enough. A BoundingBox anchors the subject. Use to gate face or identity work on frontal poses.

Segment Image Everything (SAM 3)

@meta

Prompt-free "segment everything" on an Image: a point grid segments every object with no prompt. The sink resolves the projection: mask+box pairs, masks, mask-derived boxes, parallel arrays, a merged Mask, a count, or a colourised Image. Class-agnostic.

Transcribe Audio (SenseVoice)

@funaudiollm

FunAudioLLM SenseVoice transcriber. Resamples each AudioFrame to 16 kHz, decodes it, strips the event and emotion markers, and emits a clean String (empty on silence). For timestamped output prefer transcribe_audio_faster_whisper.

Detect Turn Speech (Smart Turn v3)

@pipecat

Appends each AudioFrame to a rolling 8-second 16 kHz history and runs Smart Turn v3 over it, emitting a Double probability that the speaker's turn has ended — to gate replies on real turn-end rather than a fixed silence timer.

Flatten Audio Frames

@bleenco

Joins a [AudioFrame] batch head-to-tail into one AudioFrame, first unifying every frame to a common sample rate and channel layout so a mixed-format batch joins cleanly. Use when downstream needs one continuous segment but upstream emits a list of frames.

Detect Document Corners (DocAligner)

@docsaid

DocAligner document-corner detector: takes an Image and returns the four page corners as a Polygon quadrilateral or as [Landmark] keypoints, selected by the downstream connection. Use to isolate a page before cropping, unwarp, or OCR.

Label Halpe Body Landmarks

@bleenco

Reindexes a flat [Landmark] list in Halpe-26 order into a named Landmarks2d.Human26 record with full-body and foot fields. A Maybe<Landmarks2d.Human26> sink yields an empty result on a short list; either arm always emits exactly one output.

Pace

Regulates a fixed leg a and a variadic pack of independently typed legs to one output rate in Hz, driven by Pace's own monotonic timer. A zero-order-hold resampler: drops inputs faster than rate, duplicates the last value when idle, never interpolates.

Locate Objects (Eagle LocateAnything)

@nvidia

Takes a scene Image and a description String, runs the NVIDIA Eagle LocateAnything grounding VLM, and emits one located instance per match in pixel coordinates; the sink picks the arm (BoundingBox, Rectangle, labelled box, count, or overlay).

Inpaint Image (LaMa)

@advimman

Takes a source Image and a Mask (or second Image channel) and reconstructs the masked region with learned content; supports LaMa (fast, CPU/GPU) and SD2 (text-guided, GPU) backbones. Prefer inpaint_image_opencv for small scratches.

Output ZIP Archive File

Packs each upstream [{name: String, content: Bytes}] list into a single .zip archive under output_path, deflating members at compression_level and naming the file from name_template plus a 6-digit index.