Topaz Labs Starlight Precise 2.6 vs Starlight Fast 3: Which Video Upscaler to Use

Same price per second, one API key. The difference is speed, a sharpness control, and what you are upscaling.

Topaz Labs makes the video enhancement models behind Topaz Video and Topaz for Web, and two of them now run on the Yotta AI Gateway: Starlight Precise 2.6 and Starlight Fast 3. They are the first models in the Gateway’s new Video-Enhancement category, and they sit next to the video generation models you may already be calling, so you can generate a clip and upscale it with one key and one base URL.

Both models do the same job. You send a video, you get back a sharper, higher resolution version of it. Both are diffusion models, both go up to 4K, and both are priced identically on the Gateway. So the question is not which one is cheaper. It is which one fits the clip in front of you.

This post lays out what each model does, what a clip costs, when to pick which, and the exact request shape for both.

The short version

Fast 3 is up to 4x faster than Precise 2.6 at the same price, with what Topaz describes as comparable 4K output quality. For AI-generated clips, review copies, social cuts, and anything where you will iterate, start with Fast 3.

Precise 2.6 is the model with the Sharpen control, and it is the one Topaz built around faces, fabrics, textures, and fine text. For a final master, archival footage, or a clip where you want to dial the look, use Precise 2.6.

If you are not sure, run Fast 3 first. If the output needs more than Fast 3 gave it, run Precise 2.6 on the same input. At these prices, a second pass on a 10-second clip is under a dollar at 1080p.

What Starlight Precise 2.6 is

Starlight Precise 2.6 is Topaz’s flagship diffusion upscaler for realism. It takes footage up to 4K while rebuilding detail the source does not have, and the 2.6 release is tuned for faces, fabrics, materials, and textures, plus the clarity of text, labels, and logos. Topaz lists it as up to 15% faster than Precise 2.5, with better temporal consistency, meaning less of the periodic flicker that diffusion upscalers can introduce across long sequences.

The 2.6 release also added a Sharpen parameter. On the Gateway this shows up as a required sharpness value from 1.0 to 5.0, with 5.0 as the default. Lower values give a softer, more natural look. Higher values lean into the generative detail. That control is the main thing Precise 2.6 has that Fast 3 does not.

Topaz’s documentation positions it for footage where every frame matters and you have time to get it right.

What Starlight Fast 3 is

Starlight Fast 3 is the newest model in the Starlight Fast line, and the first Fast model Topaz describes as comparable to Precise in 4K output quality. Topaz’s numbers: up to 4x the processing speed of Precise 2.6, comparable 4K output quality, and up to 7 frames per second at 1080p, which puts a 10-second clip under 40 seconds. Topaz’s published benchmark in its own web app is a 15-second, 30 fps clip upscaled to 4K in 4 minutes 17 seconds on Fast 3 versus 18 minutes 45 seconds on Precise 2.6. Gateway times will vary with load, but the ratio is the point.

Fast 3 also adds scene splitting. It detects cuts inside a clip and treats each scene on its own, which keeps the output stable across transitions. That matters more than it sounds for AI-generated video, where a single clip often has two or three hard cuts.

Topaz specifically calls out improved skin, text, and fine detail compared to Fast 2, and the examples in Topaz’s Fast 3 documentation are AI-generated clips, including one made with Seedance 1.5 Pro. Fast 3 does not take a sharpness parameter. You get Topaz’s default look.

Precise 2.6 vs Fast 3 side by side

Speed, throughput, and quality claims are Topaz’s own. Pricing is as listed on the Yotta console at time of writing.

What a clip costs on the Gateway

Both models are billed per second of video, and the rate depends on the output resolution you ask for. There are two tiers.

Two things fall out of this table. First, 1080p costs the same as 720p, so there is no reason to stop at 720p unless your pipeline needs it. Second, 1440p costs the same as 4K, so the same logic applies at the top tier. In practice you are choosing between $0.059 and $0.124 per second, and the decision is whether you need more than 1080p.

Because Precise 2.6 and Fast 3 cost the same, speed is free. A Fast 3 pass that comes back in a fraction of the time costs exactly what the slower Precise 2.6 pass would have.

When to use Fast 3

Fast 3 is the right default for most Gateway workloads.

• Upscaling AI-generated video. Topaz’s own Fast 3 examples are GenAI clips, and the scene splitting handles the hard cuts those clips tend to have.
• Review copies and iteration. If you are going to look at the result, change the prompt, and generate again, you want the fast model in the loop.
• Social and short-form. A 10-second clip at 1080p for $0.59, back in under a minute.
• Batch jobs. When you are upscaling hundreds of clips, 4x faster is the difference between an hour and an afternoon.

When to use Precise 2.6

Precise 2.6 earns its time when the clip is the deliverable.

• Final masters. The model Topaz tuned for faces, fabrics, and fine text.
• Footage with people in it, where skin and clothing texture are what the viewer looks at.
• Archival and compressed sources, where you want the maximum rebuild.
• Anything where you want control over the look. The sharpness parameter is the only tuning knob either model exposes on the Gateway, and only Precise 2.6 has it.

A sensible pattern: Fast 3 while you iterate, Precise 2.6 once on the version you ship.

How to call either model on the Gateway

Both models use the same endpoint and the same request shape. The only differences are the model slug and the sharpness parameter.

Submit a job with a POST to the video-enhancement endpoint. The source video goes in as a single reference item. You must include the file’s size in bytes, duration in seconds, frame count, and input resolution, because Topaz uses those to validate and price the job.

Endpoint, submit:

POST https://gateway.yottalabs.ai/api/maas/video-enhancement/generations

Endpoint, poll:

GET https://gateway.yottalabs.ai/api/maas/video-enhancement/generations/{request_id}

Authentication is an X-API-KEY header with your Gateway key. A successful submit returns code 10000 and a request_id. Poll until status is completed, then read output_url. Status can also come back as failed or cancelled.

Request parameters:

• reference_items (required). Exactly one item. type is “video”. url is an HTTPS link ending in .mp4 or .mov. size is the file size in bytes. duration is the length in seconds. frame_count is the total frames. resolution is the input resolution: 720p, 1080p, 1440p, or 4k.
• aspect_ratio (required). 16:9 or 9:16. Default 16:9.
• resolution (required). Output resolution: 720p, 1080p, 1440p, or 4k. Default 1080p.
• sharpness (required for Precise 2.6 only). 1.0 to 5.0, one decimal place. Default 5.0. Fast 3 rejects this parameter.

A minimal Fast 3 request body, upscaling a 720p clip to 4K:

{"model": "slf-3", "parameters": {"reference_items": [{"type": "video", "url": "https://example.com/input.mp4", "size": 11010963, "duration": 8.05, "frame_count": 193, "resolution": "720p"}], "aspect_ratio": "16:9", "resolution": "4k"}, "providers": ["topaz"]}

The same request for Precise 2.6 changes the model to slp-2.6 and adds a sharpness value inside parameters, for example “sharpness”: 4.0.

The console model page for each model has a full Python client you can copy, with submit, poll, and a generate-and-wait wrapper. Open the API tab on the Starlight Fast 3 or Starlight Precise 2.6 page in the console to grab it.

Generate, then upscale, on one key

The reason these models are on the Gateway and not just on Topaz’s own API is the models sitting next to them. Seedance 2.5 and 2.0, Wan 3.0 and 2.7, Kling v3, and PixVerse V6 and C1 all generate video through the same base URL and the same key. That turns a two-vendor pipeline into one.

The flow is three calls. Generate a clip with a text-to-video or image-to-video model and poll for the output_url. Take that URL, plus the size, duration, and frame count of the result, and submit it to Fast 3 as the reference item. Poll again and you have a 1080p or 4K version of a clip that was generated at 720p a minute earlier.

On cost: a 10-second clip from Seedance 1.5 Pro is $0.12 at the console rate, and upscaling it to 4K with Fast 3 is $1.24. The upscale is most of the bill, which is worth knowing when you decide how many clips get the 4K treatment and how many stay at 1080p for $0.59.

If you have not used the video generation side yet, the Seedance API guide walks through the generate half of this, and the PixVerse V6 vs C1 post covers the cheapest generation models on the Gateway.

FAQ

Is Fast 3 lower quality than Precise 2.6?

Topaz describes Fast 3’s 4K output as comparable to Precise 2.6, with significantly better skin, text, and fine detail than Fast 2. Precise 2.6 is still the model Topaz tunes for faces and textures first, and it is the only one with a sharpness control. For most AI-generated clips the difference will not be the thing you notice. For a final master with people in it, run both and look.

Why do both models cost the same?

That is how they are priced on the Gateway. Topaz’s own API prices them the same way, at the same frames-per-credit rate for both models at a given resolution. The speed difference is a time saving, not a cost saving.

What input formats work?

MP4 or MOV, delivered by HTTPS URL. You also have to pass the file size, duration, frame count, and input resolution with the request. Those fields are required, not optional.

Can I upscale a video that was generated on the Gateway in the same request?

Not in one request. Generation and enhancement are separate endpoints. You poll the generation job for its output_url, then submit that URL to the enhancement endpoint. Two jobs, one key.

Are Wonder 3.5 and the other Topaz models on the Gateway too?

Wonder 3.5, Topaz’s generative image upscaler, is live in the Image-Upscale category. More Topaz video models are being integrated. Check the console catalog for what is live today.

Bottom line

Starlight Precise 2.6 and Starlight Fast 3 are the same price on the Yotta AI Gateway, so pick by the job. Fast 3 for AI-generated clips, iteration, and volume. Precise 2.6 for final masters, faces, and the one knob you get to turn. Both run on the key you already use for Seedance, Wan, Kling, and PixVerse.

Get an API key in the Yotta console and run a clip through each. The two model pages have the full request format and a working Python client ready to copy.