NightJet

Tiny luma-first low-light enhancement models for images, videos, and Jetson-class edge deployment.

Jetson Orin Nano Super	82.9 inferences/sec — 12.0 ms/frame, FP16 TensorRT @ 1280×720 (benchmark)
Model size	2,778 parameters — under 20 KB of weights

NightJet takes a short temporal window of low-light luma frames and predicts an enhanced luma frame. The public repo includes the default PyTorch weight, a matching ONNX export, examples, and the training/export tools used to produce the model.

Try NightJet live on Hugging Face. The default model artifacts are at ggamecrazy/nightjet-edge-v1. For managed Jetson/k3s deployment, use KubeJet and its EdgeVisionPipeline example.

Install

uv python install 3.12.12
uv sync --locked

Run the local checks:

just check

Enhance An Image

uv run nightjet enhance \
  --input examples/assets/input.jpg \
  --output outputs/nightjet-image.png

The default output is grayscale RGB because the model is trained on luma only. Use --preserve-color to recombine the enhanced luma with the original chroma:

uv run nightjet enhance \
  --input examples/assets/input.jpg \
  --output outputs/nightjet-color.png \
  --preserve-color

Generate a before/after strip:

uv run nightjet enhance \
  --input examples/assets/input.jpg \
  --output outputs/nightjet-comparison.jpg \
  --side-by-side

Enhance A Video

uv run nightjet enhance \
  --input path/to/low-light.mp4 \
  --output outputs/nightjet-video.mp4 \
  --preserve-color

The video path is causal: each output frame uses only the current frame and previous frames. At the start of a clip, NightJet pads the window with the first available frame.

The video path decodes variable-frame-rate sources without duplicating frames. To reduce pan ghosting, PyTorch inference and TensorRT temporal window engines drop older frames from the temporal window when cumulative block-luma motion exceeds 0.045. Tune that threshold with --motion-budget, or use --disable-motion-budget when you want to force the full causal window.

Python examples are in examples/enhance_image.py and examples/enhance_video.py.

Deploy On Jetson

Build a target-specific TensorRT engine on the Jetson/runtime image:

uv run nightjet build-engine \
  --onnx weights/nightjet-edge-v1.onnx \
  --output outputs/nightjet-edge-v1-fp16.plan \
  --fp16

Run the same image/video enhancer through a TensorRT engine:

uv run nightjet enhance \
  --input examples/assets/input.jpg \
  --output outputs/nightjet-trt.png \
  --engine outputs/nightjet-edge-v1-fp16.plan

Benchmark: Jetson Orin Nano Super

The default nightjet-edge-v1 engine, built exactly as above from the committed ONNX, measured on a Jetson Orin Nano Super Developer Kit (JetPack 6 / L4T R36.5, TensorRT 10.11, FP16, MAXN_SUPER with locked clocks, batch 1, 1280x720 five-frame luma window, trtexec sustained load, 300 iterations):

Metric	nightjet-edge-v1 @ 720p
Throughput	82.9 inferences/sec
GPU compute per frame	12.0 ms
Host→device / device→host transfer	0.84 ms / 0.25 ms
End-to-end engine latency	13.1 ms

At 12 ms per frame the model fits comfortably inside a 60 FPS frame budget (16.7 ms), so on this device the practical ceiling is the camera, not the model.

Reference Orin Snapshot

Live demo on the same device (July 6, 2026), running a heavier companion model inside a full camera pipeline: a 1280x720 camera feed enhanced in real time and served as a 3-way comparison stream (raw / classical / model) over MJPEG, capped at 15 FPS by configuration.

Runtime metric	Observed value
Displayed stream rate	15.0 FPS — the configured cap, fully saturated
Processing latency (frame read → encoded JPEG)	89.5 ms
Model GPU compute per frame	~16 ms, run in parallel with CPU work
Model cost added to the frame time	2.1 ms
Classical baseline	22.0 ms
3-way render + JPEG encode	61.5 ms (on its own thread)

A live-demo snapshot, not a model benchmark; latencies vary with scene content and exclude capture-buffer age.

For managed Orin deployments, build the runtime image from docker/Dockerfile.orin and publish it as an immutable ghcr.io/cezarc1/nightjet:<tag>-orin tag. KubeJet examples use the headless nightjet serve command; camera-specific tuning belongs in hardware-specific runtime manifests or companion repos.

Public Weights

Weight	Role	Source run	Notes
weights/nightjet-edge-v1.pt	Default	edge-v1-reco-s2-c16-f5-detail-v1-5000	Conservative 5-frame model with better temporal behavior.
weights/nightjet-edge-v1-detail.pt	Detail variant	edge-v1-reco-s2-c16-f5-reddit10-5000	Sharper, but more flicker-prone.

Matching ONNX exports and SHA-256 hashes are recorded in weights/manifest.json. TensorRT .plan and .engine files are not canonical artifacts because they are built for a specific target device, TensorRT version, precision mode, and runtime image.

See weights/README.md and MODEL_CARD.md before redistributing or using the weights beyond research/demo work.

Docs

Topic	Link
Model architecture	docs/architecture.md
Training and KubeTorch runs	docs/training.md
ONNX and TensorRT export	docs/export.md
Jetson and runtime handoff	docs/jetson.md
Hugging Face release	docs/huggingface.md
Campaign reports	docs/reports/

night-vision-orin, KubeJet, and KubeTorch are advanced companions. They are useful for camera bring-up, managed Jetson deployment, and GPU training runs, but they are not required to run nightjet enhance on an image or video.

Train Or Export

Run a CPU smoke train:

uv run nightjet train \
  --config configs/candidates/edge_v1_reco_s2_c16_f3.yaml \
  --output-dir outputs/smoke \
  --device cpu \
  --max-steps 2

Export an ONNX model:

uv run nightjet export-onnx \
  --checkpoint weights/nightjet-edge-v1.pt \
  --output outputs/nightjet-edge-v1.onnx

For full training data preparation and KubeTorch submission commands, see docs/training.md.

Limitations

• NightJet enhances luma; it is not a full RGB restoration model.
• It cannot recover scene content that the sensor did not capture.
• The detail variant can introduce visible temporal flicker.
• Evaluation so far uses small held-out clips and teacher-derived targets.
• Jetson FPS and latency claims must be measured on the Jetson Orin Nano target, not on the GPU training node.

Artifact Policy

This repo commits source code, documentation, tiny public model weights, matching ONNX files, and tiny example assets. Datasets, teacher bundles, ad hoc checkpoints, TensorRT plans/engines, and generated run outputs stay out of git.