Triage a folder in four steps
- Drop several photos, or choose a folder. Everything is decoded in this tab. Nothing is uploaded.
- Read the preflight. Every file is measured before any is treated: how many come back, how many partly, how many are past reach, how many are mosaic or already sharp — and the time it should take on this device.
- Override any row. Force a pass on a file the measurement declined, or leave a file out. Strength, guard and, if you like, one fixed kernel apply to every file.
- Treat, then download the ZIP. Files whose pass did not clear the keep-threshold go into the ZIP exactly as they arrived. A CSV report inside the ZIP says what happened to each.
The keep-threshold
Most blurred photos do not come back, and a bulk tool that hands every one of them back "enhanced" is hiding that. This one measures the detail along the axis the blur ran on, before and after each pass. A pass that did not clear the threshold is discarded and the file you gave it goes into the archive unchanged. The archive therefore only ever holds two kinds of file: a reconstruction that measurably improved on what arrived, or your original.
| Restored | {name}-deblurred.{ext} |
|---|---|
| Kept as it arrived | the original file name, the original bytes |
| Report | deblur-report.csv |
Limits
- Past-reach smears and mosaics are named, not treated, unless you force them.
- One kernel for all only makes sense for frames from one burst; otherwise let each file be measured.
- Large photos at full size can exhaust a phone's memory; pick a working size.
One photo at a time? The deblur workspace gives you the kernel dial and the loupe.
Questions
Why are some files returned unchanged?
Because the pass did not measurably help them. After each pass the detail along the blur's own axis is compared with what arrived; under the keep-threshold the reconstruction is thrown away and the original file goes into the ZIP byte for byte. Re-encoding a file for no gain would only cost quality.
What do the five verdicts mean?
Comes back: a smear along one axis, within reach. Comes back partly: missed focus, within reach, with some halo. Past reach: a smear longer than the method handles at that size. Nothing to gather: a block grid or mosaic. Already sharp: nothing to undo.
What is in the CSV report?
One row per file: its name, what it was saved as, the verdict, the kernel and method used, the edge-energy ratio, whether the archive holds the reconstruction or the original, and the bytes.
How many photos can I add?
As many as the device can hold; they are worked on one to four at a time (you choose; the default suits this device's memory). A phone does best with one or two at a time.
When won't this work?
On photos that are past reach or pixelated, which are named rather than treated; on very large photos at full size on a phone, which can run out of memory — use the 1024 or 2048 px working size.
Related
- Unblur an imageDiagnose the blur, then reconstruct it, with the kernel on a dial.
- Check if a photo is blurryMeasure blur before you upload: edge width, a heat map and a form checklist.
- Blur to clear, one tapA verdict and a result with no controls; fine-tune only if you want to.
- How it worksWhich engine ran, how each verdict is reached, and the model licence chain.
