Method

How PicRemedy cleans and upscales images locally

See the local model checks, supported inputs, background-separation steps, processing limits and why an AI-enhanced output is not recovered ground truth.

What we calculate

AI Clean runs a compact same-size neural model for noise and JPEG or WebP artifact reduction. Neural Upscale can synthesize a 2x or native 4x output after its WebGPU check passes. Compatibility resize uses ordinary browser canvas resampling and is not described as AI.

How the local queue works

  1. The browser holds up to ten selected files in page memory and accepts PNG, JPEG, WebP or AVIF files up to 30 MB each.
  2. It decodes dimensions locally, rejects work outside the active mode's envelope and processes one item at a time.
  3. AI Clean preserves dimensions. PNG stays PNG, WebP requests WebP, and JPEG or AVIF requests a JPEG result; the browser reports the actual output type.
  4. You compare the generated output and download only the selected result.

Audience measurement records only the property-free events work_started and work_completed. No file name, pixel data, dimensions, mode, scale, queue count or result value is attached.

What we deliberately leave out

Neural Upscale accepts inputs no larger than 512 by 512 pixels and 262,144 pixels. AI Clean keeps source dimensions but has a 4,000,000-pixel WebGPU ceiling and a 50,176-pixel CPU/WASM ceiling. These are admission policies, not promises of speed or visual quality. Generated pixels are model inferences, not recovered historical truth.

How to use the result

Inspect faces, text, repeated patterns and high-contrast edges at full output size. If an area looks synthetic, incorrect or over-sharp, keep the original instead.

How the background remover works

The background remover is a separate tool on this site with a separate model. It separates people from their background: the model is a person-versus-background segmenter, so it is built for portraits, group shots and full-length photos. When it finds no person it says so instead of cutting out an object, and no measurement covers other subjects.

  1. The photo is drawn into a 256 × 256 canvas in this tab and the model returns one mask value per pixel of that square. The model cost is the same for a 0.3 MP photo and a 24 MP one, because it always runs at 256 × 256.
  2. If the subject occupies less than three quarters of the frame, the tool runs a second pass on the subject's own crop, which is what puts more mask detail where the edge actually is.
  3. The mask is scaled up and refined against your photo's own luminance, which pulls the boundary back onto real edges instead of leaving the soft halo an upscaled 256-pixel mask would otherwise have.
  4. The result is composed at the source's exact pixel dimensions and encoded locally — PNG when you keep transparency, JPEG when you choose a flat colour.

What the browser checks before the control appears

The model file is downloaded once and its length and SHA-256 digest are compared against the values recorded in the release. The browser then runs a fixture built from a fixed pattern and compares the mask it produces against the mask recorded when the release was made, as a 32 × 32 grid of block means, with a maximum difference gate of 0.05 and a mean gate of 0.005. Only after both pass does the tool run your photos. A device that fails is told so; it is not offered a control that cannot work.

Why this is not face recognition

The model returns a single foreground mask covering the whole person. It does not detect, measure, template or store facial geometry, and PicRemedy computes no face landmarks, no face embedding and no faceprint anywhere in this pipeline — not in memory beyond the current operation, and not in any browser storage. This is a deliberate boundary rather than an omission: a general foreground mask is not biometric identification, and PicRemedy does not add face-aware features to this tool.

What it costs to load

The first run downloads roughly 3.9 MB compressed: a 447 KB model plus the local inference engine this site already uses for AI Clean. The page states the size before the download starts and the browser caches both files, so later visits begin immediately. Nothing about the photo leaves the tab at any point.

Where the limits are

The mask is 256 × 256 before refinement, so single strands of hair, chain-link fences seen through a gap and other detail thinner than the refinement can recover stay approximate. Ten photos fit in one queue, each up to 30 MB, and a result is admitted only inside the same working-memory ceiling the rest of the site uses. Compare at full size before you use a cut-out.

Guides that apply this method

Three longer pages work through the same limits for one job each.

Editorial policy and change log

PicRemedy describes AI output as synthesized. A generated pixel is an inference from the pixels around it, and no page describes it as the original coming back.

  • Every measured figure on a page comes from a receipt or model manifest kept with the site's source. A figure without a receipt is not published.
  • A model option appears only after its browser check passes, and each page says that a control may be absent.
  • A limit is stated next to the claim it limits, not on a separate page.
  • Nothing on the site claims that a person reviewed or tested a result.

To report a mistake, use the address on the contact page and quote the sentence. Corrections are published in the next release of the page. A change to a measured figure or a stated limit is added to the log below, newest first.

Change log

  • 13 August 2026. Background remover receipt added: the weights licence and conversion record, the browser parity check and the measured cost in four browser engines.
  • 11 August 2026. Runtime contract and AI Clean quality receipts added: input limits, model sizes and checksums, the image boundary, the analytics events and the frozen damage experiment.