5 AI Video Upscaling Routes and a Fair Test Protocol

Video upscaling can enlarge a frame, suppress noise, and invent plausible texture. It cannot recover detail that was never recorded. That distinction is the reason a polished demo is a poor buying test. The useful question is which processing route fits the source damage, privacy boundary, workstation, edit pipeline, and cost model.

This guide maps five current routes from their first-party documentation: Topaz Video, AVCLabs Video Enhancer AI, Pixop, DaVinci Resolve 21 Studio, and UniFab Video Upscaler AI. It is not a winner ranking. HUMAI did not install these applications, upload footage, buy credits, or retain comparable exports. Product behavior and image-quality claims remain unverified here.

The pages were checked on July 17, 2026. Interfaces, models, hardware requirements, and prices can change, so the final section supplies a repeatable benchmark instead of a frozen verdict. Run that benchmark on the footage you actually need to deliver before paying for a long plan or committing an archive.

Diagnose the footage before choosing an upscaler

A resolution label describes frame dimensions, not useful detail. A 720p master with mild compression may enlarge cleanly. A 1080p file copied through several social platforms can contain ringing, blocked gradients, smeared faces, and baked-in sharpening. Sending both through the same preset confuses scaling with restoration.

Start by recording the source rather than guessing from the filename. Note pixel dimensions, frame rate, scan type, codec, chroma sampling, bit depth, average and peak bitrate, and whether the frame rate is constant or variable. Keep the untouched master and a checksum. If the file is interlaced, deinterlacing belongs before a progressive upscaling judgment. If it is already over-sharpened, more edge contrast may make halos worse even when the output looks crisper at first glance.

Then mark three short segments that expose different failure modes. One should contain a face or fine texture, one should contain motion or a camera pan, and one should contain text, graphics, grain, or repeated patterns. Include a difficult transition or cut. A single flattering still can hide temporal shimmer, unstable hair, crawling edges, or a one-frame identity change.

Define the delivery target separately. A 4K file is not automatically a successful result. Specify the required dimensions, frame rate, color space, transfer function, audio handling, codec, container, bitrate range, and maximum file size. The acceptance decision should be made at normal playback speed on the intended display, with paused frames used only to locate a problem.

Five routes, five different operating constraints

The products overlap at the word upscaling, but they enter a workflow at different points. This table is a routing aid, not a quality score.

Do not force every route into one procurement category. Pixop moves footage to a cloud service and meters several parts of the job. Resolve Studio is a broader post-production purchase. Topaz documents both local and cloud execution, while AVCLabs and the UniFab route covered here are desktop applications. The right shortlist may contain only two options once data residency, operating system, and existing editor are considered.

Topaz Video is not the discontinued Video AI application

Topaz now documents two separate generations. Video AI versus Topaz Video marks Topaz Video AI as discontinued and identifies version 7.1.5 as its final release. It says no new models will be issued for that legacy application. Current Topaz Video remains in development, carries no AI in its product name, and requires an active subscription for continued access.

That product boundary matters when reading old reviews. A perpetual license, older Intel Mac support, a model list, or an interface shown for Video AI cannot be carried forward as a current Topaz Video fact. Someone who owns the legacy application can still find the entitled installer in the account's Legacy section, but a new evaluation should use the current documentation and current installer.

Topaz Video supports local processing, and its cloud rendering guide says local rendering is staying available. Cloud jobs use separately purchased credits. Cost varies with input and output resolution, frame rate, duration, and model, so there is no honest fixed price per clip. The guide also lists changing cloud limitations, including supported codecs, maximum input size, filter exclusions, and model-specific frame limits. Check that list immediately before a production upload.

The cloud route is not just a faster export button. The documentation says some Starlight use on a Personal license requires cloud rendering. It also notes that a cloud render can differ subtly from a local result and advises previewing frames before export. A private archive may therefore have two independent questions: whether its footage may leave the workstation, and whether the selected model is actually available locally under the chosen license.

Hardware is the other filter. The current Topaz Video requirements recommend modern Apple silicon or a recent Windows workstation with substantial memory and GPU capacity. Individual Starlight models have stricter, and different, requirements. Do not reduce that page to one VRAM number. Check the selected model, operating system, GPU family, driver, RAM, storage for downloaded models, and whether local processing is supported on that platform.

AVCLabs offers a local, preset-led route

The current AVCLabs Video Enhancer AI guide describes a desktop sequence: import one or more files, choose a preset or individual functions, set output parameters, preview, then process. It lists batch input, scaling from 100 to 400 percent, named targets through 8K, and MP4, MKV, MOV, WebM, and AVI output containers.

Its preview is deliberately small. The guide describes a side-by-side view of the first 30 frames, with a selectable starting point. That is useful for rejecting a bad setting quickly, but 30 frames cannot establish temporal stability over a pan, a long dissolve, or changing illumination. Use it to narrow settings, then render the full benchmark segments before judging the application.

AVCLabs documents Standard, Ultra, Standard Multi-Frame, and Ultra Multi-Frame enhancement models. The two Multi-Frame options are listed as Windows-only. Combining enhancement, interpolation, and colorization can sharply extend processing time, according to the same guide. A fair comparison should first test upscaling alone, then add one correction at a time. Otherwise, nobody can tell which operation helped, which introduced an artifact, or which caused the render-time increase.

The hardware page separates minimum from recommended configurations and publishes a specific supported-GPU list. On macOS it calls for Apple silicon. On Windows it names supported NVIDIA, AMD, and Intel routes. Compare the exact machine against the live page instead of assuming that any GPU with enough memory is accepted.

AVCLabs currently presents Video Enhancer AI as a paid desktop product on its store page. The displayed price and promotion are volatile, so record the checkout currency, tax, renewal terms, device allowance, refund rule, and update entitlement on the purchase date. This article does not turn the store's marketing description into an independent quality finding.

Pixop makes cloud cost part of the technical test

Pixop is the clearest fit when the requirement is browser or API processing without maintaining a local inference workstation. Its current resources page presents a browser path for upload, processing, comparison, and download, plus an API for integration. The service handles the compute, but the customer still has to budget transfer, storage, encoding, and processing.

The live Pixop pricing reference separates filter rates, encoder rates, storage, and downloads. Direct billing and AWS Marketplace billing follow different structures. Filter and encoder charges depend on output gigapixels, while stored and downloaded data add their own units. A rate for Super Resolution alone is not the cost of a finished delivery.

Estimate a candidate job in the live interface and save the estimate with its settings. Include every filter, the output encoder, predicted output dimensions and frame count, storage duration, download volume, tax, and the billing channel. Add reruns if the team expects more than one approval pass. Delete unused media under the retention plan rather than discovering storage charges after the project ends.

Pixop's first-party filter-testing instructions recommend making short clips, trying settings there, comparing the processed and source versions, and only then applying a preset to the full video. That sequence is especially important for metered cloud work. Build clips from several scenes, because a setting that cleans a locked-off interview can damage fast movement or fine graphics elsewhere in the same master.

A cloud route also needs an upload decision. Confirm that the organization may transfer the source, where processing and storage occur, who can access projects, how media is deleted, and what evidence is required for a client or regulator. Public security statements help frame questions, but they do not replace the contract, account configuration, or the customer's own data classification.

DaVinci Resolve 21 Studio keeps scaling inside the edit

Blackmagic Design's current product page identifies DaVinci Resolve 21 as the current generation. It offers a free edition and DaVinci Resolve Studio 21. The page places Super Scale among functions powered by the DaVinci AI Neural Engine and lists that engine with Studio, so a buyer should not assume the free download includes the same upscaling route.

The material advantage is context, not a universal claim about sharper pixels. An editor already finishing in Resolve can keep scaling, noise treatment, reframing, color management, titles, audio, and delivery in one project. That removes an intermediate file and makes it easier to compare the treated clip against the timeline around it. It also makes the benchmark responsible for the whole render chain, not only the scale control.

Test Super Scale on duplicated clips with all other corrections held constant. Inspect faces, hair, foliage, subtitles, film grain, thin diagonals, fast motion, and shot boundaries. Then return the winning candidate to the actual grade and export settings. A result that looks acceptable in the viewer can change after noise reduction, sharpening, color-space conversion, or the delivery codec.

On July 17, 2026, Blackmagic's US page displayed Studio at $295. That dated figure is useful only as a checkpoint, not a permanent global price. Check the live regional store or reseller, tax, supported operating system and GPU, license delivery, and the current manual before buying. Existing Studio owners should verify whether their installed version and project environment support the documented Resolve 21 path before migrating active work.

UniFab documents options but overstates certainty

The UniFab Video Upscaler AI manual describes four model labels: Texture Enhanced, Speed-Optimized, Quality-Optimized, and Anime-Optimized. It offers Standard and High output quality choices, resolution targets that include 1080p, 4K, 8K, and 16K, and sequential batch processing. It also says higher resolution takes longer and higher quality produces a larger file.

Those are product descriptions, not measured outcomes. The manual publishes vendor timing examples for several GPUs, but it does not make them transferable to another source, codec, driver, thermal limit, or output setting. Its Windows recommendation includes an RTX 30-series or newer GPU with 8 GB of VRAM and 16 GB of system memory. Its Mac row calls for macOS 13 or later and lists different Intel and Apple silicon configurations. Re-check the page against the exact machine.

The same manual says output quality can vary with the source's resolution and compression. Yet its answer to whether processing can cause quality loss is an absolute no. That answer should be rejected as a test conclusion. Upscaling changes pixels, resampling can alter edges, encoding can discard information, and a model can create unstable or false detail. Only a source-specific comparison can show whether the trade is acceptable.

UniFab can remain in a shortlist if its operating system, local-processing boundary, model choices, and export route fit the project. Treat 16K as an available setting, not proof that a 16K result contains meaningful new information. Benchmark at the intended delivery size first. An unnecessarily large output increases render time, storage, and encoding load while making artifacts easier to magnify.

Run one benchmark that every route must pass

Create a benchmark folder with the untouched source, its checksum, a media-information report, and three to five short lossless or source-native segments. Use identical in and out points. Include faces, motion, texture, graphics, dark gradients, and at least one difficult cut. If privacy rules prohibit cloud transfer, mark cloud candidates ineligible instead of uploading a substitute and pretending the results are comparable.

Choose one target resolution and frame rate. Disable interpolation, colorization, stabilization, HDR conversion, and creative sharpening during the first pass. Where a product requires a bundled operation, record it. Keep output codec, bit depth, chroma sampling, color tags, and bitrate as close as the routes allow. Save application version, model name and version, every slider, processing location, hardware, driver, start and finish time, upload and download time, credit use, and file size.

Review exports blind if possible. Rename them with neutral codes and randomize playback order. Score temporal stability, face consistency, edge halos, text legibility, texture plausibility, noise behavior, banding, grain continuity, motion cadence, and shot-boundary errors. Record both the affected timecode and severity. A still frame can support the note, but the decision must include full-speed playback.

Measure operations separately from picture quality. Capture active operator minutes, total elapsed time, local peak memory, GPU memory where available, cloud queue time, failed jobs, retries, billed credits, storage, downloads, and final encoded size. Price a realistic month or archive, not only the three test clips. The cheapest test can become the most expensive route if it needs repeated supervision or creates a second conform.

Use stop conditions. Reject a route if it changes identity, makes subtitles less readable, produces recurring temporal artifacts, breaks color or audio requirements, violates the data boundary, cannot reproduce settings, or has an unbounded job cost. If two routes pass, choose from workflow ownership and total cost. There is no need to invent a visual winner when both satisfy the delivery specification.

Keep the decision reproducible after the software changes

An upscaling decision expires. Save the benchmark date, source checksum, route, application version, model build, settings, hardware, processing region, license or credit terms, output checksum, review notes, and approval. Preserve a small rights-cleared test segment that can be rerun after an update. Do not keep confidential media in a cloud account merely to make future testing convenient.

Before the production batch, repeat one approved segment with the installed build and compare it to the accepted export. Pause if a model, driver, default preset, encoder, or cloud price changed. For Topaz, confirm that current Video rather than legacy Video AI is being evaluated. For Resolve, confirm Studio entitlement. For Pixop, refresh the whole-job estimate. For AVCLabs and UniFab, re-check supported hardware and platform-specific model availability.

The procurement record should name why the route was selected: local privacy, cloud capacity, editor integration, supported hardware, acceptable artifacts, predictable billing, or some combination. That reason survives a marketing page better than a numbered list. When the constraint changes, rerun the protocol and let the evidence change the choice.