Topaz Starlight Explained: Mini, Sharp, HQ, Precise, Local or Cloud, and How Much VRAM

Key takeaways

• Starlight is Topaz Video's family of diffusion models for upscaling and restoration. There are five: Mini, Sharp, Fast 2, HQ, and Precise.
• Every Starlight model needs 32 GB of system RAM on Windows, and at least 10 GB of VRAM on NVIDIA. Mini needs 20 GB on AMD.
• On a Mac, only Mini and Precise run locally. Sharp, Fast 2, and HQ are cloud-only there.
• Mini and Precise run locally on any Topaz Video license. Sharp and HQ need a Pro account to render locally.
• Topaz's own docs say to expect hours of rendering, to skip preview, and to export a few seconds as a test instead.

Topaz Starlight is a set of five diffusion models inside Topaz Video. Mini suits old, low-resolution footage, Precise suits faces and good sources, and Sharp, Fast 2, and HQ sit between them. All of them need far more memory than Topaz's classic models, and on Macs and AMD cards several run only in the cloud.

Topaz spreads these facts across its documentation and pricing pages. In this guide, you'll learn what each model is for, which ones your computer can run, what the cloud option costs you, and what to do if your machine falls short.

Contents

1. What is Topaz Starlight?
2. The five models at a glance
3. Hardware: what your computer needs
4. Which plan runs which model locally
5. Rendering locally: the rules Topaz gives
6. Rendering in the cloud
7. If your computer is below the floor
8. FAQ

What is Topaz Starlight?

Topaz Starlight is a series of diffusion-based video enhancement models in Topaz Video that rebuild detail by generating it rather than sharpening what's there. Topaz introduced the first Starlight as a cloud-only model in early 2025. Starlight Mini followed as what Topaz calls the first local desktop diffusion model for video enhancement.

Diffusion models are slower and hungrier than the classic models in Topaz Video, such as Proteus and Iris. They can produce more convincing texture on poor footage, and they also invent more of it.

The five models at a glance

Precise 2.6 replaced 2.5 in Topaz Video 1.7.0 and added three sharpness levels; level 3, the default, matches 2.5.

Hardware: what your computer needs

These are Topaz's Starlight requirements. On Windows, every model also needs Windows 10 or 11, 32 GB of system RAM, and an Intel or AMD CPU with AVX released after 2020.

A few consequences that surprise people:

• GTX 10-series and Quadro P cards are excluded from every Starlight model, whatever their VRAM, because their compute capability is too low.
• A 16 GB NVIDIA card meets the minimum, not the comfortable setting, for Fast 2. Topaz notes a slight quality difference between 16 GB and 32 GB.
• Precise targets compute capability 8.9, which means RTX 40-series and newer. Topaz says 8.6 cards (RTX 30-series) may work but aren't officially supported.
• System RAM matters as much as VRAM. A 16 GB VRAM card in a 16 GB RAM PC is below the floor.

For how VRAM and compute capability map to real cards, see which GPU you need for AI video enhancement.

Which plan runs which model locally

Topaz's pricing page sums it up for the standard plan: Precise and Mini locally, cloud access to Sharp, HQ, and Fast. For plan prices, see our Topaz Video review.

Rendering locally: the rules Topaz gives

Topaz's model pages list the same handful of rules. Following them avoids most failed Starlight renders:

1. Deinterlace first. Interlaced footage (DV, VHS, broadcast) should be deinterlaced before Starlight. See how to deinterlace video.
2. Use Starlight alone. Mini and Precise can't be combined with other filters. If you need denoise or stabilization too, export, re-import, and run Starlight as a separate pass.
3. Mind the size. Mini wants square-pixel input of 1920 × 1080 or less and outputs 1280 × 720 to 4K. Precise takes inputs up to 4K and caps output at 4K.
4. Skip preview. Topaz says preview won't work with Precise and isn't advised with Mini. Export a few seconds as a test instead.
5. Expect hours. Topaz's own Mini page says to expect hours of rendering, depending on hardware. Version 1.7.0 made Mini up to about 2× faster on NVIDIA cards with more than 12 GB of VRAM, and Precise 2.6 up to 15% faster.

If a render crashes or stalls even within these rules, our Topaz Video crash and slow-render fixes go through memory settings and logs.

Rendering in the cloud

Every Starlight model can render on Topaz's servers, which is the only route for Sharp, Fast 2, and HQ on a Mac. Three things to know:

• Cloud rendering uses credits. The Topaz Video plan includes 25 cloud credits a month. Extra credits come in one-time packs, from 20 credits for $5 to 9,000 for $999, or monthly bundles.
• The per-minute cost depends on the model and output size. Topaz's cloud page says its prices are updated periodically and may not match the app, so read the price preview in the app before you send a job.
• There are limits. Mini has a 5-minute maximum per cloud render, and Sharp's cloud output is capped at 1080p.

Cloud rendering also means uploading your footage. If that's a concern for client or family material, see local vs cloud AI video processing.

If your computer is below the floor

Many machines don't reach 32 GB of RAM and a 10 to 20 GB graphics card. You have four options:

KwaFlux is one example of the last row. It runs on Windows GPUs from 2016 or later, integrated or discrete, and on Apple Silicon Macs, with 8 GB of RAM as the minimum and 16 GB recommended. Integrated graphics work, but 4K and long clips take noticeably longer. There are no cloud credits to buy, because every job runs on your own GPU, and one license covers 2 computers and includes photo, audio, and subtitle tools as well as video. The Perpetual License lists at $199.99, paid once. Our guide to Topaz alternatives for low-end PCs compares several tools on that basis.

Here's how the hardware floor compares, using Topaz's Starlight requirements and KwaFlux's published system requirements:

The two don't run the same models, so this compares what your computer needs, not output quality. Test both on the same few seconds of your own footage.

Final thoughts

Starlight is Topaz's heaviest tool: five models, each with its own memory floor and its own local-or-cloud rule. Check the hardware table against your machine before you plan a project around it, test with a few seconds of export, and keep a lighter option for the jobs your computer can't carry.

A local option for the PC you already have

KwaFlux runs AI video upscaling and repair on Windows GPUs from 2016 or later and on Apple Silicon Macs. AI processing runs locally on your GPU, and media files are not uploaded for processing. It runs on Windows 10 (64-bit, version 1809 or later), Windows 11, and macOS 12 or newer on Apple Silicon. Previews need no account: render a 1-, 3-, or 5-second preview on any clip. Signing in with an active paid plan is required only to export, convert, repair a file, or download a file from a link.

Download KwaFlux · See plans

Related reading: KwaFlux vs Topaz Video

Topaz Labs, Topaz Video, and Starlight are trademarks of Topaz Labs. KwaFlux is an independent product and isn't affiliated with, sponsored by, or endorsed by Topaz Labs or Adobe. Requirements are quoted from Topaz's pages as of the access dates shown and may change.