AI Deblur

AI Deblur sharpens camera shake and focus miss. Two things can produce the result and the difference between them matters: an unsharp mask raises contrast either side of an edge, which reads as sharper but adds nothing that was not already there, while NAFNet is a learned deconvolution trained on pairs of blurred and sharp photographs, so it reconstructs detail the blur spread out.

The two paths

• NAFNet is a 92 MB one-time download, cached afterwards, and it runs behind two conditions: you have allowed the download, and the browser has a healthy WebGPU adapter. Both have to be true.
• The classical direction-aware unsharp mask needs no download, runs instantly, and is the path whenever either condition is not met.

WebGPU is a real requirement rather than a preference: on single-thread WASM, NAFNet crawls badly enough that it is not worth the wait. If the model is unavailable, fails to load, or throws mid-run, the classical path finishes the job — the tool always returns something, and the caption under the result names which path produced it.

Nothing downloads until you allow it

The consent gate comes before any bytes move: Allow the download fetches NAFNet once, Not now keeps the classical mask. The answer is stored on this device and can be changed later from the same line of text. See why AI tools download a model once for where the weights live and when a browser can evict them.

The blur type buttons only steer the classical path

Detect it runs a gradient analysis and picks; Motion blur sharpens across the streak; Focus miss treats the softness as uniform in every direction. All three feed the classical pipeline only — NAFNet infers the blur itself and ignores the setting. When the classical path detects directional blur it reports the angle it found next to the result, which is a useful check on whether you picked the right type.

Sizes the model will and will not accept

NAFNet's graph declares both spatial dimensions as dynamic, but it is not. Measured against the real weights, there is a hard per-axis minimum somewhere between 368 px (fails) and 372 px (passes) — below it the encoder's spatial dimensions collapse and the run aborts inside a padding node. It is per axis, not per area: 512×256 and 256×512 both fail while 480×448 is fine.

So the tool plans a working size before it runs anything. The long edge is capped at 1024 px because deblur is heavy, the short edge is lifted to at least 384 px if the cap pushed it under, and both axes are bounded at 2048 px. The result is scaled back, so what you download always has the dimensions you uploaded. The practical consequence: a very wide panorama is processed at that short-edge floor, and fine detail in an extreme crop of it is limited by that rather than by the model.

Which blur can actually be fixed

Blur destroys information, and no model recovers what was never recorded. What deconvolution can undo is a predictable spreading of light.

• Handheld shake is the best case. A short streak in one direction is what NAFNet was trained on.
• Slight focus miss is recoverable when the subject sits close to the focal plane.
• Noise mistaken for blur gets worse, not better. Deblur amplifies high-frequency detail, and low-light sensor noise is high-frequency detail.
• Heavy motion, or focus missed by a wide margin, is not recoverable here or anywhere. A tool that returns a confident result from that input is generating a plausible photo, not restoring yours.

Limits

• NAFNet needs WebGPU. Without it you get the classical mask, whatever you answered about the download.
• The model works at up to 1024 px on the long edge, so it is a detail-recovery pass, not an upscaler.
• JPEG, PNG or WebP in, up to 32 MB. PNG out.
• Deblur sharpens; it does not denoise. On a noisy source, expect the noise to come through harder.
• Everything runs in this tab. Nothing is uploaded.

In the image editor

The tool is also in the image editor's launcher, under AI Tools, as AI Deblur. It is not offered in the video editor. A run started from the editor uses the download answer you have already given; with no answer stored it takes the classical path rather than starting a 92 MB fetch from a surface that never asked.

Related

• AI Face Restore — targets faces specifically, and does more for them than a whole-image deblur can
• AI Describe — a clearer image gets a better description
• Why AI tools download a model once
• Browser support — which browsers report a usable WebGPU adapter