Topaz Labs Starlight Precise 2.5: the benchmark for AI video enhancement that actually works on hard footage

Topaz Labs has a strange advantage in AI video upscaling: it has been doing this long enough to know where enhancement models usually fail. Most new video models look good on short, clean demos. They struggle when the source is soft, compressed, noisy, old, or already generated by another AI system. Starlight 2.5 is aimed at that harder middle ground.

Topaz announced Starlight 2.5 in its March 31, 2026 Precision Update, then brought it back into the April 28, 2026 Next-Gen release as part of a broader push across Topaz Video, Astra, and its API. The March launch positioned the model around GenAI footage that needs to reach 4K without getting a plastic finish. The April launch added a second message: Topaz wants the same class of heavy video model available locally where possible, not only through cloud rendering.

That matters because Starlight 2.5 is not a small sharpening filter. It is a diffusion-based video enhancement model. It upscales, denoises, de-aliases, and sharpens while trying to keep motion stable from frame to frame. In plain terms: it is meant to make footage look as if more real visual information was available, not as if someone dragged a clarity slider too far.

What changed with Starlight 2.5

The model page from Topaz describes the target clearly: AI-generated footage with good composition and structure, but soft faces, artificial textures, or weak detail. The update also calls out text, labels, product shots, signage, faces, and modern archival footage from the 2000s onward.

That list is useful because it tells you what not to expect. Starlight 2.5 is not magic for every damaged tape in a closet. Topaz says the model is a better fit for footage at 480p and up, or footage with enough visible structure for the model to interpret. Its own desktop guidance warns against very poor source quality, heavy motion blur, interlaced content that has not been de-interlaced, and tiny or inconsistent text that the model may read as texture.

I would use it first on three kinds of clips: soft AI video that needs a 4K delivery file, older digital footage that has recognizable detail but weak resolution, and live-action shots where faces or fabric need a more natural finish. I would be more cautious with low-quality VHS, badly blurred sports footage, or documentary material where a generated texture could be mistaken for evidence.

Upscale ratios and resolution limits

Topaz's desktop documentation lists available local outputs as minimum size, 2x, 3x, and 4x. Those ratios are relative to the original video or to the model's minimum and maximum thresholds. The cap is the part to remember: Starlight 2.5 is currently limited to 4K UHD, which means 3840 by 2160, both locally and in the cloud.

Runware's API documentation exposes the same practical ceiling in a different way. You choose a target output width and height, with a maximum width of 3840 pixels and maximum height of 2160 pixels. The aspect ratio is preserved by computing the scaling from the shorter edge. Output frame rate can range up to 120 frames per second, with a floor of 15 frames per second or the input video's own frame rate, whichever is higher.

So the clean user-facing rule is simple: use Starlight 2.5 to move toward HD or 4K, but do not plan on 6K or 8K output from this model. If your clip is already 4K, the scale choice will not push it past that cap.

Pricing right now

There are two separate pricing stories.

For Topaz's own products, the pricing page currently lists Topaz Video at $299 billed annually for Personal and $699 billed annually for Pro. Topaz Studio is broader: it bundles the company's photo, video, cloud, and web apps, and currently starts at $37 per month with an annual commitment or $69 month to month. Astra, the cloud video product, is also listed separately on the pricing page, with annual Personal pricing shown as $328 billed annually and Pro as $2512 billed annually.

For the API route used by Z.Tools, pricing is per input second. Runware lists 1K upscaling at $0.080 per second at 24 frames per second, and 4K upscaling at $0.175 per second at 24 frames per second. That makes a five-second clip $0.40 at 1K or $0.875 at 4K. A 30-second clip would be $2.40 at 1K or $5.25 at 4K before any platform-specific credit packaging.

This is where the cloud wrapper makes sense. If you enhance video every week, Topaz Video or Astra may be the better buy. If you only need to test a few clips, per-second rendering avoids the subscription decision.

Generador de Videos IA

Crea videos desde texto, imagenes o transforma material existente

→