See what the dark was hiding.
Your cursor is the light. Wherever it falls, LumenX's restoration surfaces and the scene rises by the light it recovered.
Bring your own night.
Three paths, one image
The same photo through every restoration path. The best value in each row is highlighted.
Restoration report
What LumenX found
The engine's own reading of this photo, in plain English.
Look through the lens
Every map below was computed for this image. Tap one to lay it over the result.
The restoration, step by step
Tap a step to see the image at that moment.
Performance on this image
Measured before and after. Arrows show whether each change is an improvement.
Light, in three dimensions
The restored photo draped over the engine's illumination heightmap. Peaks are where the light survived; contours mark equal brightness.
Adaptive search
Every setting the engine tried for this photo, and why the others lost.
Light, measured
One line per row of the illumination map. Peaks are where light survived.
Shape the light yourself
Relight preview loads with the final image.
The room, as the camera sees it.
Frames are enhanced independently. When the lane is busy, new frames replace stale ones.
Measured, not claimed.
Every number on this page is read live from the evaluation results the engine produced. Nothing here is typed in by hand.
Reading the evaluation summary…
Head-to-head
Arrows show which direction is better. The best value in each column is highlighted.
What each part contributes
Fine-tuning
Was it worth retraining the curve model on our data?
Known limitations
Where LumenX still struggles, stated plainly.
Failure analysis
Runs the engine's own safety audit flagged during evaluation.
All datasets
Pick a dataset to see every evaluated configuration.
How LumenX sees in the dark.
Most enhancers brighten a photo and hope. LumenX measures what the camera actually recorded, restores it with a deep model, repairs what brightening breaks, and shows its working at every step.
Five steps from black to light.
Each image below is the engine's real output at that step, from the recorded run on the landing page.
- 01
Diagnose
Measures how dark the scene is, how noisy the sensor was, and how much real signal survived, before choosing anything.
- 02
Restore with CIDNet
A deep model brightens the scene in a colour space built for low light, recovering tone and colour that simple gain would wash out.
- 03
Repair highlights
Where lamps and windows would blow out, LumenX blends in a protected copy pixel by pixel instead of clipping them to white.
- 04
Clean the grain
Noise is removed only where brightening amplified it, so texture in well-lit areas is left untouched.
- 05
Audit and report
Every result is checked for blown highlights, colour shift, lost detail and halos, and each check is reported with its value, pass or flag.
Measured, not guessed
Settings are chosen per photo from what the sensor recorded, not from a fixed preset.
Shown, not hidden
Every stage, map and rejected setting is visible, so a result can be inspected instead of trusted blindly.
Honest when unsure
Unmeasured scores stay empty with a reason, and recoverability is reported as an estimate.
Built on open work.
HVI-CIDNet
Yan et al. Architecture and weights from the Fediory/HVI-CIDNet repository, MIT licence.
Zero-DCE++
Li, Guo & Loy, IEEE TPAMI 2021. CC BY-NC 4.0, non-commercial research use only.
LOL-v1
Paired low-light dataset, Wei et al., BMVC 2018. Used for evaluation only.
Unsplash & LIME
Abigail Keenan, Fons Heijnsbroek, Julien (domsson) and Julio Lopez on Unsplash; LIME benchmark images via the Zero-DCE repository.
Mona Sans · Boska
Mona Sans and Geist Mono under the SIL Open Font Licence; Boska under the ITF Free Font Licence.