The short answer on AI video upscalers: local desktop rendering wins on quality-per-dollar, cloud tools win on convenience, and the open-source stack is genuinely good if you have a Vulkan GPU and patience. For clean 1080p source going to 4K, Topaz Video on the desktop and DaVinci Resolve Studio’s Super Scale are the two options I’d put money behind, because both run on your own hardware and cost you nothing per finished minute. Cloud upscaling is not cheap: at Topaz’s own published credit rates, a minute of 4K output runs about $7 to $18.
Comparing documented limits and cost
The criteria are detail gain, temporal stability, artifact rate, documented limits and cost per finished minute. Prices are from vendor pages checked on 16 October 2026, in USD before tax unless stated. Licence terms come from the repositories.
There was no controlled side-by-side test for this comparison. Published specifications cannot establish which tool produces the best result on your footage.
What “better 4K” actually means
Real detail gain is the only metric that matters commercially, and it is the one vendors describe least precisely. Upscaling cannot recover information that was never captured; it predicts plausible substitutes. Temporal stability means detail invented identically across consecutive frames rather than shimmering.
Artifact rate is what you actually see at 200% on hair, grain and high-contrast edges. The other three are operational, and they are what break projects: documented caps on resolution, clip length and file size; cost per finished minute rather than cost per month; and licence terms that decide whether you can ship the output or the tool.
A useful framing comes from Topaz’s own published taxonomy, which is the clearest model-versus-model distinction any vendor has written down. Topaz splits its video models into precision, generative and creative. Precision models “refine existing information rather than reconstructing new detail”; generative models “may introduce inferred detail or texture not explicitly present in the source”; creative models “do not attempt to preserve original visual detail”.
That framework generalises across the whole category, and I’ll use it throughout.
The head-to-head: what each tool actually documents
This table compares only limits and prices each vendor publishes. Where a vendor states no cap, the cell says so rather than guessing. Prices were read from each vendor’s own page on 16 October 2026, in USD.
| Tool | Where it runs | Documented max output | Documented clip / file caps | Entry price, USD (checked 16 Oct 2026) | Free tier behaviour |
|---|---|---|---|---|---|
| Topaz Video (desktop) | macOS / Windows, local GPU | Capped at 16K output; H.264 High tops out at 4096×4096 | No length cap published | $299/yr annual; $59/mo monthly | Trial mode, watermark on export |
| Topaz Studio (bundle) | Local + cloud | Same as Topaz Video | Same | $399/yr annual; $69/mo monthly | Same; includes 300 cloud video credits/mo |
| Topaz for Web – Video | Browser, cloud | 1080p or 4K | Credits, not seconds; 200 credits/mo | $19/mo annual, $29/mo monthly | No free plan |
| Topaz Astra | Cloud | 4K | 400 credits ≈ 10 min of HD video (vendor’s own figure) | $19/mo annual vs $39 list | No |
| Adobe Firefly upscaler | Browser, cloud (runs Topaz Astra) | 1080p or 4K | 5–20 s, under 200MB, MP4 or MOV, one file | Firefly subscription; credits per job | Free generations, count unpublished |
| DaVinci Resolve Studio Super Scale | Local, GPU | “4K and 8K images from HD sources” | No length cap published | $295 one-time | Free Resolve has no upscaler |
| Video2X + Real-ESRGAN | Local, Vulkan GPU | Whatever your GPU and encoder allow | No vendor cap | $0 | Fully free, no watermark |
| NVIDIA RTX Video SDK | Local, RTX 20-series or newer | 1080p “with 4K precision” detail | SDK-level | Free SDK | N/A: SDK, not an app |
Two things jump out. DaVinci Resolve Studio is the only one-time purchase in the professional tier, and the free edition of Resolve simply has no upscaler: Super Scale sits behind the DaVinci Neural Engine, which Blackmagic lists among the Resolve FX you only get in Studio. Adobe’s Firefly cap is brutal and completely documented: twenty seconds per upload.
Decision one: what does a finished 4K minute actually cost?
Local rendering costs GPU time, not money, which is why the desktop tools win on this metric by an order of magnitude. Topaz Studio includes “unlimited local rendering”, and Topaz’s single-app pages make the same promise for desktop purchases.
Cloud is where real money goes, and Topaz Labs publishes the per-render rates directly in its Cloud Apps documentation:
“Starlight Precise 2.6: 5 credits to 1080p, 20 credits to 4K” for a 10-second clip at 30 FPS: Topaz Labs, Plans & Pricing | Cloud Apps
Do the arithmetic, because nobody does it for you. Twenty credits per ten seconds is 2 credits per second, so 120 credits per finished 4K minute.
Topaz sells one-time credit packs from $3.00 for 20 credits ($0.15 each) down to $180 for 3,000 credits ($0.06 each). That puts a finished 4K minute at $18.00 at the small-pack rate and $7.20 at the volume rate. The subscription does not rescue you: the $19/month Video plan carries 200 credits, which is 1 minute and 40 seconds of 4K output per month.
Runway’s developer API is a useful independent anchor, and it prices upscale per output frame rather than per second. Its published table lists 720p at $0.0072 per frame, 2K at $0.009 and 4K at $0.012: 360 credits for a ten-second 30 fps clip at 4K. At $0.01 per credit that is $21.60 per finished 4K minute.
| Route | Basis (source) | Cost per finished 4K minute, USD |
|---|---|---|
| Topaz Video / Resolve Studio, local render | Unlimited local rendering (Topaz pricing page) | $0 in software; GPU hours only |
| Topaz cloud, 3,000-credit pack | $0.06/credit × 120 credits | $7.20 |
| Topaz cloud, 20-credit pack | $0.15/credit × 120 credits | $18.00 |
| Topaz for Web – Video, $19/mo plan | 200 credits/mo ÷ 120 credits/min | ~$11.40 |
| Runway Dev API, 4K upscale | $0.012/frame × 1,800 frames | $21.60 |
| Video2X + Real-ESRGAN | BSD/AGPL software | $0 in software |
Two caveats on that table. Topaz notes its rates are estimates based on 10s at 1080p 30fps, and that higher frame rates consume credits “proportionally based on total processed frames”, so 60 fps roughly doubles it. And for HD rather than 4K output the same Topaz model costs 5 credits per ten seconds, a quarter as much.
Decision two: what happens when the footage can’t leave the building
If your material is unreleased, client-confidential, or archival, local rendering stops being a preference and becomes the only answer. Topaz makes the capability claim directly: precision models “are well suited for local workflows”, while generative diffusion models “are significantly larger and more computationally intensive” and, “with limited exceptions such as Starlight Mini and Starlight Sharp, most generative video models require cloud processing”.
That’s a real constraint, not marketing. Diffusion-based upscaling at 4K is a heavyweight job, and if your only option is the most aggressive model, you may be forced to upload. Video2X and Real-ESRGAN are entirely local by construction, and Resolve Studio is local too.
HitPaw VikPea is the interesting middle case: it documents local processing and cloud rendering as separate options, with a published four-GPU render at 18 minutes against two GPUs at 2 hours 38 minutes and one GPU at 3 hours 16 minutes.
Decision three: caps, watermarks and batch reality
Check the caps before you upload, not after. These are published, and they decide which tool fits which job.
- Adobe Firefly: 5 to 20 seconds, under 200MB, MP4 or MOV, one file at a time. All four limits appear on Adobe’s own upscaler page.
- Topaz for Web: metered in credits, not seconds, so the limit is your wallet rather than a hard stop.
- Filmora: the free edition adds a watermark to every export, per Wondershare’s own FAQ. AI credits are issued every 30 days and “any unused credits from the previous 30-day period are reclaimed”: no rollover.
- TensorPix: the free plan caps input at 720p, holds 1GB, and deletes files 30 days after creation. Paid plans raise input to 4K. Output codecs are H.264 and H.265 across every tier, with one concurrent job on Free rising to seven on Elite.
- Batch: TensorPix publishes concurrency explicitly: 1, 3, 5 and 7 concurrent enhancements across Free, Standard, Premium and Elite. HitPaw VikPea documents batch upscaling but does not publish a job cap.
For agencies running a queue, the number that matters is concurrency and storage, not the headline model count. And check how you are allowed to pay: Remaker AI sells credits only, one-time, with no subscription, and its own FAQ states that it accepts crypto and Telegram payments while card payments “are not supported and won’t be available in the near future”. That is a procurement conversation, not a rounding error.
Decision four: where AI helps and where it invents
Precision models refine; generative models reconstruct; creative models reinterpret. That distinction is the single most useful thing any vendor has published about this category, and it comes from Topaz Labs.
Topaz classifies Proteus, Proteus Natural, Iris, Dione, Artemis, Gaia, Rhea, Rhea XL and Theia as precision video models; Starlight Precise 1 and 2, Starlight Fast 1 and 2, Starlight Mini and Starlight Sharp as generative; and Astra Creative as creative. Topaz’s own words on the generative family: “the diffusion process may infer detail or texture that is not explicitly present in the original footage.”
That maps cleanly onto your use case:
- Documentary, news, unscripted, archival: precision, locally, with human review. Topaz states precision models are the “preferred first approach” for nonfiction.
- AI-generated or synthetic source: generative, because the source detail doesn’t exist to be preserved.
- Stylised or social content: creative, where you’re explicitly not trying to preserve anything.
Adobe draws the same line in its own UI. Firefly’s upscaler exposes Precise mode, which “focuses on accurate restoration of original footage”, against Creative mode, which “adds richer detail for more visual impact, especially useful for AI-generated or stylized clips.”
And Adobe’s help documentation is refreshingly blunt about where its generative tools break. On which clips work best, Adobe says clips with less contextual change, such as a fast-sweeping camera movement, perform well, and that “clips that do not work well yet include archival clips with a lot of film grain or other noise.” If you are an archivist reading this, that is the most important sentence on the page.
TensorPix is equally candid in its own FAQ: “sources with severe compression artifacts or significant quality issues may not yield substantial improvements,” while moderately degraded sources “are more likely to show significant enhancement”. Trust the vendor who tells you when not to bother.
What hallucinated detail actually looks like
Inferred detail fails in predictable places, and if you know where to look you can catch it in ten seconds. It concentrates on high-frequency edges and thin structures, because that is what the model is guessing hardest at: hair against a bright background, eyelashes, chain-link fencing, brickwork, text on a sign, a face turned three-quarters away.
It also fails temporally. A single frame can look perfect and still be wrong, because the invented detail changes on the next frame. Scrub rather than play, if a face holds still at 24 fps and then swims when you step through frame by frame, you have found invented detail.
My rule after watching this category mature: if the source is noisy, denoise first and upscale second, or pick a precision model that refuses to invent anything. Grain gives a generative model something plausible to invent with, and it will take the invitation.
Decision five: the licence line you’ll actually cross
Real-ESRGAN is BSD 3-Clause, Real-CUGAN ships under MIT, and Video2X is AGPL-3.0.
| Component | Licence (from the repository’s own licence file) | Commercial implication |
|---|---|---|
| Real-ESRGAN | BSD 3-Clause, © 2021 Xintao Wang | Commercial use permitted; retain copyright notice and disclaimer |
| Real-CUGAN | MIT, © 2022 bilibili | Commercial use permitted; retain the notice |
| Video2X | GNU AGPL version 3 | Network copyleft applies to modified versions offered over a network |
| Topaz Video, Resolve Studio, Filmora, HitPaw | Proprietary EULAs | Output terms per vendor; check the EULA, not the marketing page |
This matters more than it first appears. Running Video2X yourself to make renders is ordinary use; wrapping a modified Video2X in a hosted upscaling service and letting users reach it over a network is precisely the situation AGPL was written for. I’m not your lawyer and this is not legal advice, but if you are an agency considering a self-hosted upscaling product, get a real opinion on it.
Video2X’s own README lists its dependency licences, which is good practice and worth reading: FFmpeg (LGPLv2.1/GPLv2), ncnn (BSD 3-Clause), Anime4K (MIT), realcugan-ncnn-vulkan (MIT), rife-ncnn-vulkan (MIT) and Real-ESRGAN-ncnn-vulkan (MIT).
Upscaling is not enhancement, and conflating them costs you money
Most “video enhancers” bundle five different operations, and only one of them is upscaling. Denoise, deblur, frame interpolation, HDR restore and object removal each change the picture in ways that can undo the value of a good upscale.
- Denoise is a separate decision. Topaz lists Nyx as a dedicated high-resolution denoise model rather than folding denoising into its upscalers, and NVIDIA’s SDK treats artifact reduction as its own effect. Denoise in your NLE first if the source is grainy; sharpening noise harder just makes it sparkle.
- Deinterlace before you upscale. VHS, Hi8 and MiniDV are interlaced, and comb lines become permanent the moment they’re enlarged. Topaz’s own taxonomy lists Dione among its precision video models, separate from the upscalers.
- Frame interpolation adds frames, it doesn’t add resolution. Topaz lists frame interpolation as a distinct function on both its desktop app and its web plans; HitPaw lists a Frame Interpolation Model; Video2X wraps RIFE for the same job. Runway’s developer API makes the distinction explicit by billing “Enhance Frame Rate” separately from video upscale.
- SDR-to-HDR is a tone-mapping decision, not a detail decision. Topaz lists SDR to HDR and Stabilisation as separate functions; NVIDIA’s RTX Video SDK treats super resolution, artifact reduction and HDR tone mapping as three distinct effects; Runway’s developer API bills SDR-to-HDR separately from upscale.
- Object removal is an editorial decision, not a technical one. Blackmagic describes the Resolve FX plug-in as one “used to remove entire objects from a scene”.
The plumbing: audio, hardware and APIs
Audio is where cloud tools quietly break your timeline, and it’s the most under-documented part of this whole category. Topaz Video is the clearest because it exposes the choice: select “Copy” to retain the original audio streams “as long as the output container is compatible”, or “Convert” to re-encode. When they conflict, Topaz displays a message explaining which audio codecs are incompatible with the container.
TensorPix’s troubleshooting is the more useful warning. Their documentation notes that specific codec settings inside supported formats can be incompatible, and that “removing the audio from the video before enhancing it often resolves this”. If you care about your mix, strip the audio, upscale the picture, relink.
On hardware, the requirements diverge sharply. Topaz publishes minimum and recommended specs: macOS 13 minimum, 16GB RAM, 45GB storage; 32GB RAM and 60GB recommended, and notes separately that “the Starlight Series exceeds the app’s recommended specifications”, with its own requirements calling for macOS 15 and 36GB RAM.
Video2X needs far less: an AVX2 CPU and a Vulkan-capable GPU, from NVIDIA Kepler (GTX 600 series) or AMD GCN 1.0 upward. NVIDIA’s RTX Video SDK requires a GeForce RTX 20-series or newer, and HitPaw VikPea lists 8GB RAM minimum on Windows and macOS 13, recommending 16GB.
On APIs, the picture is good and improving. Topaz publishes per-use API pricing in credits per ten seconds at 1080p 30fps, with families for Precision Upscale, Generative Upscale (Fast and Quality), Creative Upscale, Denoise, Motion and Video Utilities. Runway charges per output frame for upscale and per second of input for frame-rate conversion.
TensorPix and HitPaw both document API access, with TensorPix listing self-hosted infrastructure as an Enterprise-tier feature.
Why your 4K upscale might not survive the upload
Upscaling to 4K does not help if the platform re-encodes it away, and YouTube’s own upload settings set the ceiling. They put 4K SDR at 35–45 Mbps for standard frame rates and 53–68 Mbps for high frame rates, against 8 Mbps for 1080p and 80–160 Mbps for 8K.
Read that again. A 4K upload wants roughly five times the bitrate of a 1080p upload. If you upscale 1080p to 4K and then encode at anything near 8 Mbps, you have thrown away the upscaled detail before it left your machine.
YouTube also notes that to view new 4K uploads in 4K you need a browser or device supporting VP9: worth remembering when a client reports that “the 4K file looks worse”. It may be the delivery path, not your master.
The seven-step pipeline I’d actually run
Most bad upscale results come from skipping the first four steps. This is the order I use, and every step is documented method rather than a vendor workflow.
- Measure the source honestly. Check the real coded resolution, the bitrate, whether it is interlaced, and how much noise is baked in. A file labelled 1080p with a 900 kbps bitrate has no detail to recover, whatever the tool promises.
- Deinterlace before anything else. Upscaling comb lines makes them permanent. Blackmagic lists de-interlacing among the Resolve FX you only get in DaVinci Resolve Studio, alongside SuperScale; Topaz keeps its deinterlacer as a separate model in its own taxonomy.
- Denoise in your NLE, conservatively, and check it frame by frame. Over-denoising before an upscale gives you plastic skin and flat skies that no upscaler can undo.
- Choose the model class before the tool. Precision for anything factual or archival. Generative only when precision genuinely fails. This is the decision that matters most and the one most comparisons skip.
- Render one ten-second test and inspect it at 200%. Check hair edges, flat skies, text on screen and faces. Scrub through it rather than watching it: flicker hides in playback.
- Export to an intermediate, not the delivery codec. Topaz Video’s encoder list includes FFV1, which its documentation describes as a “lossless compressed codec intended for use in archival workflows”, alongside ProRes and DNxHR. Grade and finish from the intermediate.
- Encode the 4K master to the destination’s published bitrate, then check it back. If it is 4K for a platform, the rate is roughly 35–45 Mbps SDR at standard frame rates.
Video2X ships a public standard test clip with a ground-truth 1080p version specifically so you can benchmark your own setup. Use it. Most people’s first assumption about their GPU’s throughput is wrong.
Gaps to check before buying
Adobe Creative Cloud and Filmora prices remain unconfirmed here. TensorPix’s cited pricing and help pages do not establish how many enhanced-video minutes one credit buys, so do not budget around the “two minutes per credit” figure repeated elsewhere. VSDream’s availability is also unconfirmed.
There is no independent head-to-head quality benchmark cited here. Test the tools on your footage before choosing one.
The Real-ESRGAN and Real-CUGAN update notes cited for architecture and licensing date from 2021–2022. YouTube’s encoding settings and FFmpeg’s scaler documentation are living documents; check them before export.
Frequently asked questions
Which AI video upscaler produces the best 4K quality right now? On documented capability, Topaz Video on the desktop and DaVinci Resolve Studio’s Super Scale lead, because both expose multiple models and render locally so your GPU is the only ceiling. Nobody has published an independent side-by-side benchmark, so treat ranked winner lists accordingly.
What is the real cost per finished minute? Local rendering is free in software terms. Cloud runs from about $7 to $22 per finished 4K minute depending on vendor and pack size, using each vendor’s published rates.
Which one is free and runs on my own machine? Video2X with Real-ESRGAN or Real-CUGAN. It needs an AVX2 CPU and any Vulkan GPU from roughly 2012 onward, and it costs nothing.
Why does upscaled video flicker? Because many upscalers process each frame as an independent still image and invent slightly different detail on each one, so that detail never sits still. The fix is a model built to look at neighbouring frames together, which is exactly what the research literature means by temporal consistency, and why metrics such as tOF were invented to measure it.
Can DaVinci Resolve upscale video to 4K? DaVinci Resolve Studio includes Super Scale, which Blackmagic describes as creating 4K and 8K images from HD sources. Studio costs $295 as a one-time licence. Super Scale requires the paid Studio edition.
Does Adobe Premiere Pro include an AI video upscaler? The Premiere Pro 26.5 release notes cited here do not list a detail-generating upscaler. Adobe’s dedicated option is Firefly’s video upscaler, which uses Topaz Astra and limits uploads to 20 seconds and 200MB.
Can I use the results commercially? Real-ESRGAN (BSD-3) and Real-CUGAN (MIT), yes. Video2X is AGPL-3.0, which mainly constrains how you distribute a modified Video2X rather than what you do with the rendered video, but get legal advice before shipping a hosted service.
Which is the cheapest free option inside an editor? Filmora’s free tier watermarks every export, per Wondershare’s own FAQ, so it is only free if you are genuinely fine with that. HitPaw VikPea’s free trial and TensorPix’s free plan both cap what you can push through (720p input on TensorPix) so “free” means proof of concept, not delivery.
Does upscaling help if I’m only posting to YouTube? Sometimes. YouTube re-encodes everything, so a clean 4K master at 35–45 Mbps gives the platform more to work with than a compressed 1080p file. It does not give you detail that was never captured.
What about archive footage? Use a precision model, run it locally, and keep the unprocessed master. Adobe’s own documentation notes that clips with heavy film grain are among those that generative tools do not yet handle well.
The bottom line
If I were buying one thing today, it would be DaVinci Resolve Studio at $295, not because Super Scale beats every dedicated tool, but because you need an NLE anyway, and the upscaler comes with it. If upscaling is the bulk of your work, Topaz Studio at $399 per year is cheaper than buying Topaz Video alone at $299 per year plus whatever else you need, and it includes unlimited local rendering plus 300 cloud credits a month.
For short AI-generated clips, Adobe Firefly’s upscaler is the same Topaz Astra model inside a tool you may already pay for, capped at twenty seconds and 200MB. For archives and client-confidential work, the open-source stack is the honest answer, and Real-ESRGAN’s BSD licence means you can build on it commercially without asking anyone.
And do the test clip. Ten seconds, 200% zoom, scrub it. That will tell you more than any comparison table, including this one.
Sources
All URLs below were opened and read on 16 October 2026 unless otherwise noted.
Topaz Labs
- https://www.topazlabs.com/pricing
- https://www.topazlabs.com/topaz-video
- https://www.topazlabs.com/astra
- https://www.topazlabs.com/learn/responsible-nonfiction-ai-enhancement (published February 2026; used for the precision/generative/creative taxonomy)
- https://docs.topazlabs.com/topaz-video/system-requirements
- https://docs.topazlabs.com/topaz-video/reference-guide
- https://docs.topazlabs.com/topaz-video/reference-guide/supported-file-formats-encoders-and-containers
- https://docs.topazlabs.com/cloud-apps/topaz-for-web-plans-and-pricing
- https://docs.topazlabs.com/cloud-apps/astra/plans-and-credit-system
- https://developer.topazlabs.com/getting-started/model-pricing
Blackmagic Design
Adobe
- https://www.adobe.com/products/firefly/features/video-upscaler.html
- https://helpx.adobe.com/premiere/desktop/whats-new/whats-new.html (version 26.5, September 2026)
- https://helpx.adobe.com/premiere/desktop/edit-projects/edit-with-generative-ai/generative-extend-faq.html (last updated 25 May 2026)
- https://helpx.adobe.com/creative-cloud/apps/generative-ai/generative-credits-faq.html
Open-source stack
- https://github.com/k4yt3x/video2x
- https://github.com/k4yt3x/video2x/releases
- https://raw.githubusercontent.com/k4yt3x/video2x/master/LICENSE
- https://docs.video2x.org/
- https://github.com/xinntao/Real-ESRGAN (repository last updated 2021–2022; used for model variants and inference behaviour)
- https://raw.githubusercontent.com/xinntao/Real-ESRGAN/master/LICENSE
- https://github.com/bilibili/ailab/blob/main/Real-CUGAN/README_EN.md (change log ends February 2022)
- https://raw.githubusercontent.com/bilibili/ailab/main/Real-CUGAN/LICENSE
Commercial and cloud tools
- https://www.tensorpix.ai/pricing
- https://help.tensorpix.ai/en/articles/8282490-faq
- https://docs.tensorpix.ai/essentials/credits
- https://www.hitpaw.com/vikpea-video-enhancer.html
- https://www.hitpaw.com/
- https://unifab.ai/unifab.htm
- https://filmora.wondershare.com/shop/buy/buy-video-editor-na.html
- https://remaker.ai/pricing/ (rendered in-browser; credit-pack and payment-method terms verified)
- https://developer.nvidia.com/rtx-video-sdk/getting-started
- https://docs.dev.runwayml.com/guides/pricing/
Standards, delivery and research
- https://support.google.com/youtube/answer/1722171 (undated living document)
- https://ffmpeg.org/ffmpeg-scaler.html (living document)
- https://arxiv.org/abs/2103.14006 (Real-ESRGAN degradation model, ICCV 2021)
- https://arxiv.org/abs/1811.09393 (TecoGAN, temporal coherence metrics, 2018–2020)
Competitor pages reviewed for this guide
