Komponenten

Translate Audio (SeamlessM4T v2)

@meta

Downmixes one AudioFrame to mono, resamples to 16 kHz, and runs SeamlessM4T v2 speech-to-text translation into a configured target language, returning a single String; silent or failing audio yields the empty string.

Estimate Metric Depth (Depth Pro)

@apple

Zero-shot metric monocular depth (Apple Depth Pro): one forward over an Image yields a deploy-time-resolved sharp MetricDepthMap in metres or its greyscale Image visualisation, with no camera metadata needed upstream.

Merge Point Clouds

@bleenco

Row-stacks every cloud in a [PointCloud] batch into one combined PointCloud, carrying optional colours and normals through when present; it concatenates only, performing no spatial alignment or deduplication.

Extract Text (PaddleOCR)

@paddlepaddle

Runs PaddleOCR (PP-OCRv5) detection and recognition over each Image, emitting either a joined String transcript or per-line regions with Polygon or Rectangle bounds and optional confidence. Optional orientation and unwarping.

Output PDF HTTP

Renders a structured Document — or an incrementally accumulating Document.Page stream — into a PDF and serves it live over HTTP and WebSocket, optionally POSTing every rendered document to a destination of its own. Each block is drawn inside the box it occupies on its source page, so the rendered PDF keeps the page's layout.

Fuse

Multi-rate re-clocker aligning several independently-rated input legs onto one common timeline with passthrough typing: output port i carries the exact type of leg i, so a a leg stays a. Faster legs pass 1:1, slower legs are held, one snapshot emitted per tick.

Reconstruct Scene (Pi3)

@yyfz

Permutation-equivariant feed-forward 3D reconstruction from a set of [Image] views with Pi3. One reference-free joint pass regresses per-view [CameraPose] and [PointMap3D], or a single fused PointCloud, with no privileged view ordering.

Decompose Intrinsic Image (Marigold)

@prs_eth

Marigold diffusion intrinsic-image decomposer: takes an Image and emits a single Image layer — the view-independent albedo (reflectance) or the packed BRDF material map (roughness in red, metallicity in green) — at the input resolution.

Sketch

Undeployable placeholder for sketching backend intent before a real component exists. Variadic [$a] inputs and [$b] outputs use unrelated typevars; a readme config key holds the intent. Strict inference refuses any backend with it.

Calculate Minimum Distance

@bleenco

Computes inter-object distances between tracked objects in the 2D image plane: takes a per-frame list of tracked-object records and emits, per object, the minimum Double distance to each other class — or the single nearest neighbour — as a list of distance records.

Output Mesh File

Exports each upstream Mesh to a 3D geometry file under output_path, choosing the container from format (obj, stl, ply, gltf, glb). Each call writes its own file named from name_template plus a 6-digit index.

Repeat

Takes a count UInt64 on the first input and a value t on the second, and emits that value n times each paired tick via repeat. Output type is exactly t; a count of 0 emits nothing that tick.

Output PDF File

Writes an upstream Document to a paginated PDF under output_path. Each block is drawn inside the box it occupies on its source page, over that page's raster, so the written PDF keeps the page's layout. A Document.Page stream accumulates into one growing PDF rewritten as pages arrive.

Calibrate Camera (Depth)

@bleenco

Calibrates a camera online from each Image: finds orthogonal vanishing points from scene line structure to recover focal length and principal point, optionally fusing known-size [BoundingBox] and a MetricDepthMap into a running CameraIntrinsics.

Output JSONL File

Writes each upstream String (or every element of a [String]) as one trimmed JSON-Lines record under output_path. Append mode accumulates records in one file; rotate mode writes each record to its own file with a 6-digit index suffix.

Detect Document Layout (PP DocLayout)

@paddlepaddle

Detects layout regions on each page Image with PP-DocLayout (PaddleX) and emits one detection per region as geometry-only [Rectangle] or [BoundingBox] carrying the model's 23-category layout class id and confidence, fixed at deploy by the consumer.

Input JSONL File

Streams records from JSONL, NDJSON, or JSON-array files as individual typed t values at a paced rate, looping the directory when asked. Use it when downstream needs typed records from a JSON corpus.

Answer Table Question (HuggingFace)

@huggingface

Answers a natural-language question about a table with HuggingFace TAPAS: takes a JSON table String and a question String, selects the answer cells and an aggregation operator in one pass, and emits the answer String (SUM / AVERAGE / COUNT computed, NONE joined).