Consent before inference
NAFNet is never requested in the background. The tool stages the operation and waits for the model decision.
Free local tool
AI image deblur
Use learned deconvolution for recoverable camera shake and mild focus miss, or an instant full-resolution classical pass when the model is declined or unavailable. The result identifies which path actually ran and stays separate from the source.
or drop one here — JPEG, PNG or WebP, up to 32 MB. It never leaves this device.
Before you download NAFNet deblur
Checking what this device can do with NAFNet deblur…
One-time download
NAFNet is a real learned deconvolution — it reconstructs detail rather than just raising local contrast, which is what separates a recovered photo from a crunchy one. It is a 92 MB one-time download, cached afterwards, and it needs WebGPU to run at a usable speed.
If you decline: the direction-aware unsharp mask runs instead — instant, no download, and genuinely useful on mild shake, but it cannot invent detail the sensor never recorded.
Choose an image first.
Result proof
A sharper file with its method attached
01
Input
Soft source photo
02
Local pass
NAFNet or classical pass
03
Result
Same-size PNG
How to verify it: Inspect text, eyelashes, rooflines and other strong edges for ringing or invented texture; the result label names the method that produced those pixels.
Blur-aware enhancement
Reconstruction and sharpening are different jobs
The optional model attempts to reconstruct detail lost to focus or motion; the local fallback increases edge contrast without claiming to reverse the blur. Both return an inspectable canvas edit with undo history.
Before / after
Review contract
Fallback stays honest
Direction-aware sharpening can improve legibility immediately, but its result is labelled as sharpening rather than model deconvolution.
Edge artefacts stay visible
A full-size preview makes ringing around text, hair, and high-contrast borders easy to reject or undo.
Know before you process: No deblur method can reliably recreate text, faces, or texture that the source contains no usable signal for.
Which blur can actually be fixed
Blur averages neighbouring detail. A short, predictable camera movement can leave enough evidence to work backwards; severe subject motion or a badly missed focal plane cannot. The tool is designed to make that distinction visible instead of turning every input into a confidently sharpened claim.
What people use this for
Short handheld shake
The strongest reconstruction case: one short blur direction across most of the frame.
Slight focus miss
Recover local contrast when the subject landed close to the focal plane, then inspect fine edges.
Soft scans and compressed copies
Use the classical full-resolution path when the problem is mild and preserving every source pixel matters most.
Preparation before upscaling
Resolve recoverable blur before adding pixels, so the upscaler does not faithfully enlarge the smear.
Limits worth knowing
- Heavy motion and severe focus miss discarded information that no deblur model can reliably recover.
- The learned path works at a bounded internal size even though the downloaded PNG keeps the source dimensions.
- Noise is high-frequency content and can be amplified into grain or false detail.
- WebGPU and device memory affect whether NAFNet is practical; the fallback remains available and is labelled honestly.
How the local workflow runs
- Load the best available original rather than a screenshot of it.
- Review the readiness and model-consent state before starting.
- Run the pass; cancellation leaves the source unchanged.
- Compare important edges at 100% and reject ringing or invented detail.
- Download the separate PNG or continue into the editor.
Frequently asked questions
Can every blurry photograph be recovered?
No. Short camera shake and slight focus miss retain useful signal. Heavy subject motion, severe focus miss and clipped detail are gone; a sharp-looking invention is not recovery.
What is the difference between NAFNet and the fallback?
NAFNet performs learned deconvolution after consent. The fallback raises local edge contrast at full resolution and does not claim to reconstruct missing pixels.
Why does the model use a smaller working size?
Browser memory and inference time are bounded. The result keeps the source dimensions, but the learned reconstruction is limited by the model working size disclosed in the tool.
Should I denoise or deblur first?
Strong sensor noise is amplified by sharpening and deconvolution. If noise dominates, reduce it first; otherwise deblur before upscaling so the blur is not enlarged.
Does deblur upload my image?
No. The image stays in the tab. If you approve NAFNet, only its model files are downloaded and cached locally.
What should I inspect before downloading?
Check high-contrast edges for halos, repeated texture, altered text and artificial detail. Compare at 100%, not only in the fitted preview.
Related tools
Ready to try it?
No upload, no account, and the result always names its processing path.
Deblur a photo →