Topaz Video Upscale Generative: restoring footage with Starlight
Topaz Video Upscale Generative rebuilds old and AI-made video with Starlight diffusion models. A tester's notes on the five engines, one trick and its limits.
I have a folder of family clips from an old camcorder, the kind where a birthday looks like it was filmed through a steamed window. I wanted to know whether Topaz Video Upscale Generative was a gimmick or the real repair. These are my notes after trying it, plus the one habit that saved me the most time.
What makes it different
The classic upscalers sharpen and enlarge what is in the frame. This one restores and upscales real and AI-generated video with the Starlight diffusion models, which means it can rebuild detail the camera barely caught. That is both its strength and its risk, and I will come back to the risk.
Inputs and the five engines
You give it one video. There are two advanced settings: the engine and the target resolution (720p, 1080p, 2k or 4k, default 1080p, aspect ratio preserved). No prompt. The engine list has five names:
- Starlight Precise 2.6, the default: high-quality generative upscaling and enhancement up to 4K.
- Starlight Precise 2.5: improved realism for AI-generated video.
- Starlight Fast 3: the fastest of the family.
- Starlight Fast 2: fast diffusion upscaling with higher fidelity.
- Starlight Mini: diffusion restoration for archival and degraded footage.
What I actually saw
On the camcorder birthday, Mini was the one made for this job and it showed: the smeared tablecloth got a pattern again and the candles stopped being orange blobs. The default Precise engine on the same clip was cleaner but a bit more confident about things that were not there. On a clip made by an AI video model, where the original problem was a mushy, over-smooth look, Precise 2.5 gave the most natural result and the best skin and fabric.
A small decision list
- Archive tape or a degraded phone file: Mini.
- Fresh AI-generated video that looks plastic: Precise 2.5.
- Best general quality when you are not sure: Precise 2.6.
- Lots of clips and a deadline: Fast 3, then Fast 2 if you want more fidelity.
Run it
In the studio choose the model, upload the clip and press Generate; the price shows on the button first. Keep the resolution modest for tests. If the clip is clean and only needs a plain enlargement or one named fix, the engine menu of the standard Topaz video upscaler may be the better tool. If the footage is AI-made and you are still choosing a generator, choosing an AI video model is a good start. Compare the neighbours on the models page.
The honest limitation
A diffusion model invents plausible detail. In my birthday clip a relative's face came back crisp and slightly different from the real one. For family memories I accepted that; for anything that must be exact, such as evidence, a number plate or a signature, do not. Fast engines also trade some fidelity for speed, and a long clip at the highest resolution is the slowest and priciest thing you can ask for. Cut first, test, then commit.
Questions
Which engine is the default?
Starlight Precise 2.6, a high-quality generative upscaler and enhancer. The other four are Precise 2.5, Fast 3, Fast 2 and Mini.
Does it work on AI-generated video as well as real footage?
Yes, it is described as restoring and upscaling both. Precise 2.5 is the engine aimed at more realistic AI-made video.
Can it change how people look?
It can. A diffusion model rebuilds detail by estimating it, so faces and small text may differ slightly from the original. Compare with the source before you rely on it.
