Komponenten

Serve SQLite Database

@sqlite

Exposes a local SQLite table as a paginated, read-only HTTP JSON endpoint via serve_sqlite_database. A GET returns recent rows newest-first with offset, limit and time-window parameters; an optional bearer token guards data requests.

Output Image File

Encodes Image or VideoFrame inputs to PNG or JPEG and writes them to disk under a prefixed, zero-padded index, optionally subsampling the stream and capping storage to a rolling window of files.

Understand Chart (ChartGemma)

@ahmed_masry

Takes a chart Image and a question String, runs a PaliGemma-based vision-language pass over the pixels, and emits a single answer String reading values, comparing series, or summarising the trend.

Deblur Image (FFTformer)

@kkkls

Frequency-domain transformer that removes camera / subject motion blur from an Image, running the GoPro/HIDE checkpoint and emitting a sharper RGB Image at the original resolution. Use for dynamic-scene motion blur, not noise or defocus.

Stamp Message

Wall-clock timestamper. On each tick it reads the host wall clock and emits a nanosecond-precision UInt64 Unix timestamp alongside the forwarded t input, both locked to the same tick so the stamp and the value stay aligned downstream.

Translate Audio (Canary)

@nvidia

NVIDIA Canary v2 speech-to-text translator. Resamples each AudioFrame to 16 kHz mono and runs the model between a configured source and target language, emitting the translated String (empty on silence). Covers a European language set under CC-BY-4.0 weights.

Translate Text (NLLB)

@meta

Runs Meta NLLB-200 sequence-to-sequence translation over a String or [String] between two explicit FLORES-200 codes, forcing the target token at decode, and returns a string, string list, or language-tagged record chosen from the resolved downstream port.

Track Objects (BoostTrack)

@pipelogic

BoostTrack multi-object tracker over per-frame [BoundingBox] or [Segmentation]. A SORT-family Kalman tracker that boosts detection confidence and inflates soft-BIoU overlap before association, so weak detections in crowded scenes survive and keep their ids. Class-aware gating gives stable ids. Emits one active-track list per frame, not per box.

Estimate Metric Depth (Depth Anything V2)

@bytedance

Metric monocular depth from Depth Anything V2: turns each Image into a deploy-time-resolved greyscale depth Image or a MetricDepthMap in metres (near small, far large), capped to a fine-tuned maximum range.

Colorize Image (DDColor)

@alibaba

Colorizes a grayscale or faded Image with the DDColor dual-decoder model loaded offline from a bound weights bundle, emitting a colorized Image while preserving the original luminance.

Estimate Face Age and Gender (AuraFace v1)

@fal

Per-face age and binary-gender estimator over an Image plus [Rectangle] face boxes, running an ONNX network per crop and emitting one box-ordered record on the resolved arm: age+gender objects (default), gender strings, age ints, a count, or an overlay.

Answer Document Question (Pix2Struct)

@google

Visual document question answering with Pix2Struct: takes a document Image and a natural-language String question, renders the question onto the page and decodes a free-text String answer in one pass, without a separate OCR stage.

Check Object Distance

@bleenco

Proximity flag for a pair of tracked objects: takes two class ids as Int64 and their separation as a Double and emits a Bool that is true when the distance is within the configured per-class-pair threshold.

Estimate Surface Normals (Marigold)

@prs_eth

Diffusion-based monocular surface-normal estimator (Marigold) over each Image, emitting a deploy-resolved choice of a dense per-pixel SurfaceNormalMap (camera-space unit normals) or a normals-to-RGB visualisation Image.

Extract Text (GOT OCR)

@weihaoran

Transcribes an Image to a String with the GOT-OCR2 vision-language model, reading the whole page or an interactive region in plain or layout-preserving markdown/LaTeX form. Suited to dense documents, formulas, and tables.

Classify Image (Triton)

@nvidia

Letterboxes and normalizes each Image, runs it through a Triton-hosted model, and projects one internal ranked result into the connected output arm: top-1 Classification, ranked [Classification], a String label, and more.

Detect Objects (DEIMv2)

@intellindust

Runs a DEIMv2 real-time DETR detector on each Image, pulling a COCO checkpoint from the HuggingFace hub at startup; the connected consumer selects the output shape, defaulting to [BoundingBox].

Output JSON HTTP

Serialises each upstream value of generic type t to JSON and fans it out over the channels in transports: a WebSocket broadcast on port 8999, an SSE stream on the same path, and/or an HTTP POST to url, each channel independent.