Diana — YOLO26 Object Detection In Pure Rust
Diana is YOLO in Rust: YOLO26 on Candle from real Ultralytics, 3.71× less CPU, a third the memory, ByteTrack, WASM — the ffai-diana crate in FFAI.
YOLO became the default "what is in this frame" stack the week it was good enough. Doorbells, shop floors, meeting rooms — if you wanted boxes without a research team, you reached for Ultralytics and PyTorch. That is a mini data center on a single board. ONNX Runtime is the other compromise. Neither is a mesh node story.
Diana is YOLO in Rust: YOLO26 on Candle, from real Ultralytics weights, matching PyTorch at 3.71× less CPU and a third the memory. ByteTrack for tracking. WASM. No Python, no ONNX, no vendored weights. The crate name in FFAI is ffai-diana. The catalog is Remade with Rust. The thesis is rebuilding the foundation of the internet. This article is the detector.
Why Diana Is YOLO In Rust Without The Python Tax
A detector that needs Ultralytics plus PyTorch is not edge AI. It is a GPU rental with extra steps. The NIST Secure Software Development Framework pushes memory-safe languages for the parsers and runtimes around models. Diana is that push for boxes. YOLO in Rust means the hot path is Candle + Rust, not a CPython wheel.
ffai-diana is how FFAI exposes Diana. Same engine, media-library surface. Pull Diana when detection is the product. Pull FFAI when detection sits next to OCR and ASR.
Trust, Security, and Incentive fail when "we detect locally" still means a Python sidecar. Trust fails when the frame left for a hosted vision API. Security fails when the runtime is a pile of native wheels. Incentive fails when the safer path misses mAP or needs a 4090. Diana is how those three primitives apply to a bounding box.
What Diana Ships As ffai-diana
github.com/Remade-With-Rust/diana is the source. YOLO26, from real Ultralytics — not a toy architecture with a familiar name. Weights are downloaded, not vendored, licenses surfaced. That split is honest. Diana is YOLO in Rust. The weights keep their licenses.
Diana Matches PyTorch At 3.71× Less CPU
The measured claim: Diana matches PyTorch at 3.71× less CPU and a third the memory. That is the Incentive. YOLO in Rust that loses on mAP does not ship. This one matches and wins on the bill. ffai-diana is the crate you depend on when FFAI is already in the graph.
Read the number as a snapshot. Workloads move. The direction does not: Diana is the detector you run when you refused Python.
ByteTrack And WASM In Diana
ByteTrack is cross-frame tracking — the person in frame 12 is the person in frame 13. A doorbell without tracking is a flicker of boxes. Diana ships it. WASM is how YOLO in Rust runs in a tab. That is the same honesty test as the rest of Remade with Rust: no C toolchain, no ONNX Runtime.
AI on the edge is the pipeline article. Diana is the "see" stage. remade_ffmpeg_rs opens the frame. Diana boxes what matters. Carmenta reads the text. Mercury hears the room.
No Python, No ONNX, No Vendored Weights
Those three refusals are the product. Diana is not a wrapper around Ultralytics. ffai-diana is not an ONNX session with a Rust handshake. Weights are not smuggled into the crate to dodge a license page. The Electronic Frontier Foundation has documented how "on-device vision" claims collapse when the helper daemon still uploads the frame. Diana keeps the frame.
The FTC keeps catching products that collected more than they needed "to detect a person." YOLO in Rust on the device deletes that excuse.
How Diana Feeds The Distributed Cloud
A distributed cloud that uploads every frame to "detect" is a surveillance vendor. Diana is YOLO in Rust so the node sees.
Trust The Frame, Secure The Detector
Trust is a frame that never left the camera you hold. Security is Diana — ffai-diana on Candle, memory-safe, no Python attack surface. Incentive is matched mAP at 3.71× less CPU and a third the memory. The NIST Zero Trust Architecture tells operators to stop trusting the perimeter. Diana is how you stop trusting the vision API as well.
The Freedom Guide is local-first steps. A camera in that guide that still shells out to PyTorch is a joke. Diana is how the joke ends.
Disco Party And Era 4 Cameras
Disco Party is the mesh. Era 4 is a home companion with a camera that answers to you. That companion runs Diana or it streams the living room. Era 1 stores the event, not the raw feed, if the box already ran on-device.
Rust on Arduino is the board story. Diana is the model story. Together they are YOLO in Rust next to the sensor.
The Catalog Entry For Diana
Source is github.com/Remade-With-Rust/diana. The FFAI surface is ffai-diana. Cards and shipping status live on the Remade with Rust catalog. The mesh this detector feeds is what a distributed cloud is and the Disco Party. Digital Freedom is Trust, Security, and Incentive.
If you came from the Learn index looking for YOLO in Rust, this is Diana. If you came from Ultralytics, this is the remake — same family of weights, no Python. The internet you want to own still has cameras. Diana is how that sentence stays true without the cloud GPU.
The reader beside this detector is Carmenta. The media pipeline under both is remade_ffmpeg_rs. The person-facing version is local AI without uploading your life.