Free AI Image Upscaling: 4 Routes, One Fair Test

If the image is private or the job contains many files, begin with a local route. Upscayl supplies a desktop interface, while Real-ESRGAN and waifu2x-ncnn-vulkan expose more control through local commands. If installation is the blocker and cloud processing is acceptable, Adobe Firefly offers a browser route with an account and a changing free allowance.

That is a route choice, not a quality ranking. HUMAI did not retain the input files, exported results, settings, or run logs behind the old version of this article. It would be false to call these tools tested here or to name a winner. This guide instead maps current documented boundaries and gives you a common-input test that produces evidence you can inspect.

The scope is narrow: increasing the pixel dimensions of a raster image and deciding whether the output artifacts are acceptable. Image generation, object removal, photo animation, video processing, and general editing suites are outside the comparison.

Define the deliverable before touching a model

Write down the required pixel dimensions and the final use. A 1200 by 800 image enlarged to 2400 by 1600 has a 2x scale on each axis and four times as many pixels. That arithmetic says nothing about whether a shoe texture, face, logo edge, or tiny letter is now more truthful. An upscaler estimates missing detail. It cannot recover information that was never captured.

Set a stop condition beside the dimensions. For a product page, it may be clean silhouette edges at the site's display size. For a print, it may be no visible halo at the expected viewing distance. For scanned art, it may be continuous line work without added speckles. If exact documentary detail matters, an invented eyelash or altered character is a failure even when the result looks sharp.

Also name the data boundary. A family photo, unreleased product image, client asset, or identity document should not be placed in a web service merely because the first generation is free. Decide whether cloud upload is permitted before comparing convenience.

Four free routes, with the limits exposed

The table reflects official pages checked on July 17, 2026. It does not promise that every operating system, GPU, model, or free allowance will remain available. Open the live release or account screen before a batch. A code license also does not grant rights to an input image, settle a bundled model's terms, or make a transformed output safe for every commercial use.

Upscayl is the local GUI route

Upscayl's official repository describes a free, open-source desktop app for Linux, macOS, and Windows. Its backend uses Real-ESRGAN with Vulkan. The same page warns that a Vulkan-compatible GPU is required and that many integrated GPUs do not work. Compatibility is therefore the first test, not an assumed benefit of installing the app.

The desktop route is useful when a person wants visible model and output controls without uploading the source to an upscaling website. This does not prove that the whole application makes no network requests. Updates, optional statistics, downloads, and organizational software controls are separate questions. For regulated or confidential work, inspect the build, settings, network policy, and approved installer rather than treating the word local as a security review.

The project's release notes document PNG, JPG, and WebP output handling, custom width, tile settings, and scales above 4x that use a 4x model followed by resizing. That last distinction matters. A 16x file is not evidence of four rounds of recovered detail. For a fair trial, record the app version, selected model, scale, compression setting, tile setting, source hash, and output hash. Reject the route if it crashes, produces seams, drops transparency, or changes a critical edge.

Real-ESRGAN exposes the local pipeline

Real-ESRGAN's official project page offers a portable NCNN executable and a Python route. The portable package includes binaries and models, requires no CUDA or PyTorch setup, accepts a file or directory, and documents 2x, 3x, and 4x output for JPG, PNG, or WebP. The Python script adds options such as arbitrary final scale, tiling, half or full precision, and an optional face-enhancement stage.

More controls create a better audit trail, not an automatic quality advantage. Save the exact command and model file checksum. If face enhancement is enabled, treat it as a separate intervention because it may change identity-bearing detail. If the Python route applies a 4x model and then a conventional resize to reach an arbitrary scale, describe both stages. Do not label the final dimensions as model-native detail.

The repository also warns that the portable implementation may introduce inconsistent blocks because it splits an image into tiles and stitches them together. Look for seams in skies, walls, gradients, and repeating fabric. A smaller tile may solve a memory failure while making boundaries easier to reveal. That tradeoff belongs in the result sheet. The CLI route fits a repeatable batch only after the same command succeeds on representative files and the output metadata is checked.

waifu2x-ncnn-vulkan narrows the art case

The waifu2x-ncnn-vulkan repository provides portable executables for Windows, Linux, and macOS across Intel, AMD, NVIDIA, and Apple silicon graphics paths. It accepts individual JPG, PNG, or WebP files and directories, exposes denoise levels, and writes PNG, JPG, or WebP. The upstream waifu2x project began with anime-style art and also says it supports photos; the current NCNN package includes separate art and photo model folders. Input and model pairing therefore belong in the trial.

Use the art model as a separate hypothesis when the source contains flat colors, inked outlines, manga tones, or deliberate digital edges. For a photograph, start with the documented photo model rather than treating the route as art-only. Keep denoise off for the first pass unless noise removal is part of the stated job. Strong denoise can erase intentional grain or thin marks, so it must not be smuggled into an upscaling comparison as if only resolution changed.

The command exposes scale values up to 32x. Large dimensions can be generated, but the number alone does not validate detail. Test the smallest scale that reaches the deliverable. Inspect diagonal lines, repeated hatching, speech-bubble text, transparency, and color boundaries. If a model turns a rough pencil line into a smooth vector-like stroke, decide from the art brief whether that is correction or damage.

Adobe Firefly trades installation for an account

Adobe's current Upscale images guide lists Firefly Upscaler, Topaz Bloom, and Topaz Gigapixel in the model menu, with 2x and 4x choices. It says credit consumption varies by model and directs the user to the indicator beside the Upscale button. Adobe's public image upscaler page requires an Adobe ID and accepts JPEG, PNG, or WebP files from 512 by 512 pixels up to 100MB. The upload screen also asks the visitor to confirm they are at least 18.

The same page says the upscale opens as a separate image that can be downloaded, but it does not freeze the downloaded file format. Confirm the live export format, color profile, transparency, and dimensions with one disposable file before the real asset. Do not infer that a PNG upload must return a PNG with the same metadata.

Do not freeze a free generation count into a workflow. Adobe's current credit FAQ says free users receive limited daily generations on a curated set of models, and that models, features, and limits may change. Some models may consume different amounts. Check the live model menu, expected dimensions, and credit indicator before each trial. If the needed 4x route is unavailable to the free account, record unavailable rather than substituting a different feature.

This is a cloud workflow. Adobe says generated media is automatically saved in Firefly generation history, consumes Adobe Cloud storage, and can be permanently deleted from that view. Adobe's general terms say cloud content may be automatically scanned for illegal or abusive material, that users retain ownership, and that customer content is not used to train generative AI models unless submitted to Adobe Stock. The terms also note that backup copies may remain for a period after deletion. Those rules are more specific than a casual claim that an upload is private.

Build one packet that every route receives

A useful packet has three rights-cleared raster files, not twelve unrelated examples. Include one compressed photograph with skin, hair, fabric, and a smooth background; one illustration with fine lines, flat color, and transparency; and one interface crop with small text and one-pixel rules. Remove names and confidential material. Keep the untouched originals in a read-only folder and record each SHA-256 hash, dimensions, color profile, alpha channel, format, and byte size.

Choose one target that all shortlisted routes can produce. A 4x target works for the documented paths above, but use it only when the deliverable needs those dimensions. Save every result under a neutral code such as A, B, C, and D. Do not put the tool name in the review filename. Record the tool version, model, scale, denoise or face option, compression, processing location, account state, credits consumed, elapsed time, failures, and output hash in a separate key.

Run each file once with default restoration settings, then allow one documented adjustment if the first output fails. That limit prevents endless tuning for a favorite route. Do not compare a denoised output with an untouched output without marking the extra operation. Do not give one route a clean PNG while another receives a heavily compressed derivative. The same source bytes must cross every starting line.

If you have a high-resolution original, add a controlled reference test. Downsample a copy with one recorded method, upscale that derivative, and compare it with the held original. This reveals invented edges and lost texture more clearly than judging a file whose true detail is unknown. The project image still matters because a laboratory crop cannot represent every defect in the real asset.

Inspect the failure, not the wow factor

Review outputs at 100 percent zoom first, then at the actual display or print size. At 100 percent, scan a fixed set of crops: eyes and teeth, fabric weave, foliage, text, diagonal edges, transparent borders, gradients, and repeated patterns. Look for ringing, crunchy oversharpening, plastic skin, doubled lines, false letters, color shifts, tile seams, checkerboard texture, and halos against a flat background.

At delivery size, ask a simpler question: does the output solve the stated problem without introducing a visible lie? Sharp but incorrect text is worse than soft text. A plausible eye that changes a person's expression is not restoration. Added sneaker weave may be harmless for a mood board and unacceptable for a product record. Artifact tolerance comes from the job, not the model's marketing category.

Score each crop with a short ordinal scale: 0 for unusable, 1 for distracting damage, 2 for acceptable at delivery size, and 3 for acceptable at both delivery size and 100 percent. Keep written reasons beside scores. Add hard checks for dimensions, readable file, alpha preservation, profile handling, and target byte size. A route fails if any non-negotiable check fails, regardless of its average aesthetic score.

Blind the first visual pass when possible. Reveal the route key only after the notes are locked. Then compare cost boundary, privacy, setup, repeatability, and correction time. This prevents a familiar logo or a complicated command from receiving unearned credit.

Make the choice reversible

Choose the route that passes the artifact threshold and the operating boundary with the least unresolved risk. For sensitive or repeated work, that may be a reviewed local installation. For a single public asset on unsupported hardware, it may be the browser route while its free allowance exists. For drawn art, it may be the specialized local model only if the line character survives the blind review.

Keep the source, route key, commands or settings, output, and rejection notes together. Recheck a cloud plan before another batch and recheck a local release before updating it. If no output passes, stop. Use the smaller image, rescan or re-export the source, commission a redraw, or change the layout. More pixels are easy to produce. Defensible detail is the actual deliverable.