Upscale Video with AI in 2026: Open Source, Desktop, or API

The cost to upscale video with AI in 2026 ranges from zero to several dollars a minute, and the spread has less to do with quality than with how you run the model. The same 60-second 480p clip costs $0.12 to enhance at 720p through an API, $2.64 through a dedicated upscaling model, or nothing at all on your own GPU with open-source software. A desktop subscription sits in the middle at $59 a month.

I pulled current numbers for all three routes: open-source tools (video2x, Real-ESRGAN), the Topaz desktop subscription, and per-second cloud APIs. This guide covers what each one costs, where each one breaks down, the steps that stay the same regardless of tool, and a Python script you can run against an upscaling API today.

TL;DR

• Open source is free and genuinely good. video2x (20,913 GitHub stars, AGPL-3.0) wraps video super-resolution and frame interpolation behind one interface^[2]; Real-ESRGAN (36,472 stars, BSD-3-Clause) is the restoration model most other tools build on^[3]. You pay in GPU hardware and setup time instead of dollars.
• Topaz Video, the desktop benchmark, is $59/month on the personal plan with unlimited local rendering and 25 cloud credits; commercial use is limited to orgs under $1M annual revenue^[1].
• APIs bill per second of source footage. Topaz Video Upscaler through reAPI runs $0.044/s on the standard tier and $0.077/s on max^[4]; the lighter enhance-video-1.0 starts at $0.002054/s for 720p output^[5].
• The break-even against the subscription is about 22 minutes of footage a month at the $0.044/s rate. Process less than that and the API is cheaper; process much more, every month, and the desktop license wins.
• The honest caveat: open source assumes you have VRAM to spare and tolerance for Python environments. If either is missing, the free route is the most expensive one in hours.

Three ways to upscale video with AI

Every AI video upscaling workflow lands in one of three buckets, and picking the bucket matters more than picking the tool inside it.

The routes also differ in what they optimize for. Local tools give you unlimited reruns and total privacy. The desktop app gives you a tuned UI and pro-grade models without any code. The API route gives you scale: a render farm you never maintain, priced per second, that slots into a script or a product backend.

Open-source upscalers: video2x and Real-ESRGAN

If your footage stays on your machine and your GPU has headroom, open source is hard to argue with. video2x is a machine-learning video super-resolution and frame interpolation framework with 20,913 GitHub stars, released under AGPL-3.0^[2]. It handles the full pipeline: extract frames, upscale each one, interpolate where needed, reassemble with audio.

Real-ESRGAN is the model doing the heavy lifting in much of this ecosystem. The project describes itself as practical algorithms for general image and video restoration, carries 36,472 stars, and ships under the permissive BSD-3-Clause license^[3]. It was built for real-world degraded footage rather than lab benchmarks, which is why it holds up on compressed, blocky source material.

Two things to check before committing. First, the license: AGPL-3.0 on video2x has implications if you embed it in a service you sell, while Real-ESRGAN's BSD license does not. Second, the hardware bill is real. Upscaling runs frame by frame, so a long clip at 4× occupies a consumer GPU for hours, and low-VRAM cards force tiling that slows things further. We keep a separate breakdown of which open-weight video models fit which GPU in our local video model guide.

The desktop route: Topaz Video at $59 a month

Topaz Video is the reference desktop product for this category. The personal plan is listed at $59/month and includes the video models, unlimited local rendering, and 25 monthly cloud credits; the full Topaz Studio bundle is $69/month^[1]. Both plans permit commercial use only for organizations under $1M in annual revenue, so agencies and larger teams need a different license^[1].

The subscription makes sense under a specific usage pattern: a person at a workstation, restoring or upscaling footage every week, who wants sliders and previews instead of flags and scripts. Unlimited local rendering means the marginal clip is free, which is exactly what you want when a restoration project takes forty attempts to dial in.

What the desktop route cannot do is run inside your product. There is no Topaz desktop API to call from a backend, and render time is your machine's problem. That is the gap the per-second services fill.

Per-second APIs: what AI video upscaling costs per clip

API pricing for video enhancement bills on the length of the source clip, which makes costs easy to predict: probe the duration, multiply by the rate. Two models on reAPI cover the range.

Topaz Video Upscaler takes a video_url and an upscale_factor of 1×, 2×, or 4×, pushing resolution toward 4K. It bills per second of source footage: $0.044/s on the standard tier, $0.077/s on max^[4].

enhance-video-1.0 is the budget tier and covers more than resolution: upscaling, denoising, deblocking, smart frame interpolation, and color refinement, with five scene presets including one tuned for AI-generated footage^[5]. Standard-tier rates run $0.002054/s for 720p output up to $0.016429/s for 4K^[5].

What a one-minute clip actually costs:

The break-even math against the desktop subscription: $59 buys about 1,340 seconds, or roughly 22 minutes of footage, at the $0.044/s Topaz API rate. Below that monthly volume, pay per clip. Above it, and if the workload is steady, the flat license wins on price. For bulk cleanup jobs where cinema-grade restoration is overkill, the arithmetic tilts hard toward the light model: ten minutes of 1080p enhancement costs $2.46, less than a single minute through the heavy upscaler.

Credits on reAPI are prepaid at 1 credit = $0.001, and failed tasks refund automatically, so a botched source URL costs nothing^[4].

Steps to enhance video resolution with AI

The workflow is the same five steps whether you run local software or call an API.

1. Assess the source. Resolution, compression artifacts, noise, and frame rate decide the treatment. Clean 1080p going to 4K is a different job than blocky 480p screen recording.
2. Pick scale before tool. Upscaling 2× is usually safe; 4× from very low resolution invents detail and can look synthetic on faces. If the source is merely compressed rather than small, denoising and deblocking at native resolution often beats resolution gain.
3. Run a short test segment. Ten seconds tells you almost everything. Locally this saves GPU hours; on an API it costs a cent or two at the light tiers.
4. Process the full clip. Batch by episode or shot list. On the API route this is a loop over URLs; locally it is a queue you babysit.
5. Verify on the target screen. An upscale that looks sharp on a phone can shimmer on a TV. Check motion-heavy sections, since interpolation and temporal artifacts show up there first.

Upscale video with Python in one script

The API route needs about twenty lines. Submit the job, poll until it finishes, download the result:

import requests, time

API_KEY = "rk_live_..."  # create at reapi.ai/settings/apikeys
HEADERS = {"Authorization": f"Bearer {API_KEY}"}

task = requests.post(
    "https://reapi.ai/api/v1/videos/generations",
    headers=HEADERS,
    json={
        "model": "topaz-video-upscaler",
        "video_url": "https://your-cdn.com/source-480p.mp4",
        "upscale_factor": "2",
    },
    timeout=30,
).json()

while True:
    poll = requests.get(
        f"https://reapi.ai/api/v1/tasks/{task['id']}", headers=HEADERS
    ).json()
    if poll["status"] in ("completed", "failed"):
        break
    time.sleep(2)

print(poll["output"] if poll["status"] == "completed" else poll["error"])

Swap the model for enhance-video-1.0 and add "tool_version": "standard", a "scene" preset, and a target "resolution" to run the cheaper enhancement pipeline instead^[5]. Input must be a public HTTPS URL; base64 uploads are not accepted.

FAQ

What are the best tools for AI video enhancement?

It depends on where the work runs. On your own GPU, video2x and Real-ESRGAN are the established open-source stack^[2][3]. For a no-code desktop workflow, Topaz Video is the category benchmark at $59/month^[1]. For automation or product integration, a per-second API such as Topaz Video Upscaler or enhance-video-1.0 on reAPI covers both quality tiers^[4][5].

Can I upscale videos using open-source software?

Yes. video2x packages super-resolution and frame interpolation into one AGPL-3.0 framework, and Real-ESRGAN provides BSD-licensed restoration models that many other tools embed^[2][3]. The practical requirements are a CUDA-capable GPU, patience for frame-by-frame processing, and attention to the AGPL terms if the tool ends up inside something you sell.

What are the steps to enhance video resolution with AI?

Assess the source footage, choose a scale factor that matches its condition, test on a ten-second segment, process the full clip, and review the result on the screen size it will actually play on. The tool changes; the sequence does not.

Do I need a GPU to upscale video with AI?

Only for the local route. Open-source upscalers run frame-by-frame inference and want a recent NVIDIA card with generous VRAM. Desktop apps also lean on local hardware. API-based upscaling needs no hardware at all, which is the point: the render happens server-side and you pay per second of footage^[4].

How much does it cost to upscale a one-minute video?

Through an API: from $0.12 for standard 720p enhancement to $4.62 for max-tier dedicated upscaling^[4][5]. On a desktop subscription the marginal clip is free after $59/month^[1]. Locally it costs electricity and GPU time.

Does AI upscaling add real detail or just sharpen?

Modern models generate plausible detail rather than stretching pixels; Real-ESRGAN was specifically trained for restoring degraded real-world footage^[3]. The generated detail is convincing on textures and scenery but can drift on faces and text at aggressive scale factors, which is why a 2× pass is the safer default.

Is 4K upscaling worth it for short-form social clips?

Usually not. Feeds compress uploads aggressively and most viewing happens on phones. Enhancing to 1080p at $0.25 a minute covers the visible difference for a fraction of the 4K cost^[5]; save 4K passes for footage headed to a TV screen or archive.

Picking a route in practice

Match the route to the volume and the destination. A one-off restoration project fits the open-source stack if the GPU already exists. A weekly editing habit justifies the Topaz subscription, since unlimited local rendering rewards iteration. Anything automated, bursty, or built into a product belongs on the API route, where a month of doing nothing costs nothing and heavy months scale without new hardware. Run the 22-minute break-even against your own footage volume before paying for either subscription or credits. However you upscale video with AI, test ten seconds before committing an hour of footage; every route charges for confidence, and the test is the cheapest thing on the bill.

References

1. Topaz Labs. Topaz Video — plans and pricing. Retrieved August 2026 from topazlabs.com/topaz-video
2. k4yt3x. video2x — machine learning-based video super resolution and frame interpolation framework. Retrieved August 2026 from github.com/k4yt3x/video2x
3. Xintao Wang et al. Real-ESRGAN — practical algorithms for general image/video restoration. Retrieved August 2026 from github.com/xinntao/Real-ESRGAN
4. reAPI. Topaz Video Upscaler — model page and live per-second pricing. Retrieved August 2026 from reapi.ai/models/topaz-video-upscaler
5. reAPI. Enhance Video 1.0 — model page and live per-second pricing. Retrieved August 2026 from reapi.ai/models/enhance-video-1-0

Further reading

• reAPI. Topaz Video Upscaler API docs — request schema and examples. reapi.ai/docs/topaz-video-upscaler
• reAPI. enhance-video-1.0 API docs — tiers, scene presets, and resolutions. reapi.ai/docs/enhance-video-1-0