Project Starlight - Video AI 6.1 Beta 3

6.1.0.5.b update


This updated beta build enables up to 9000 frames of Starlight rendering using cloud credits. We are currently pricing these renders under cost to allow as many users as possible to experience Starlight early. We will have more news on access to Starlight very soon!


Original 6.1 Beta announcement copied below:

Today’s release is a very important moment in the history of Topaz Video AI. We are releasing Project Starlight, the first-ever diffusion model for video enhancement.

This is a significant evolution over our existing models, and it’s the single largest increase in model capability since Video AI launched.

Even the most challenging footage can be enhanced with Project Starlight. So please, try out the “bad” footage—footage that may have previously produced artifacts with current models.

Because of its size, this model is significantly slower and more expensive to run than previous models, and requires cloud processing on server-grade graphics hardware. But this is just the beginning. Our mission as a product and research organization is to advance visual quality to the furthest extent possible—it just takes a little more processing power to get there right now.

We see a clear path towards expanding access to Starlight and other models of this generation to very high end desktop GPUs, and Project Starlight is just the start of a new series of models.


Project Starlight Research Preview

Active Video AI users can currently process three 10-second weekly previews of Starlight for free. Results will be viewable through unlisted, shareable links with the new “Compare” button in Video AI.

These previews take about 20 minutes to render if servers are available immediately, and all exports are set to 1080p.


Starlight Early Access

Video AI users can also render up to 9000 frames of video using Starlight (5 minutes at 30 fps).

The current introductory pricing of 90 credits/minute will change over time as our team expands access and server capacity.


Our research team is eager to receive feedback and discuss this step-change in AI video upscaling. We appreciate your involvement on this journey to Video AI’s next frontier.




Project Starlight announcement on X

Project Starlight is now also available in the Video AI web app.

6.1.0.6.b

  • Fix encoder preventing certain Starlight exports from being successful
  • Improve trim accuracy for certain sources for Starlight
  • Auto pop-up cloud exports after Starlight export

6 Likes

Question about the button that says “Render 5 min.” If your source file is under 5 mins, is it still going to use 5 mins worth of credits, or does it count by frames, seconds, or…?

The frame estimate will respond to the trim length selected.

“Render 5 min” will adjust to the framerate of your input. For example, 24 fps inputs can actually render up to 6.5 minutes. We’ll be adjusting the placement of the labels and buttons for the next build, along with some other improvements like audio in exports.

3 Likes

Neither the Workspace folder bug nor the Rhea XL download bug I reported on in the previous beta version have been fixed in this version. Please fix these bugs before releasing v.6.1

2 Likes

To be clear, nothing has changed in how Starlight processes movies? (Fine details, facial expressions or the like.)
This update just enables longer video clips, right?

1 Like

Starlight looks at hundreds of frames in order to restore a single frame. Longer video will potentially help with the quality. But how far it can go really depends on the input quality.

4 Likes

So if I double the frame rate prior to sending it to Starlight that should give Starlight more material to work with and increase the quality of the output. Is this correct?

No, that is not additional data. Those interpolated frames are created from the original source, so no actual new information is visible to the model.

Also, doubling the framerate would reduce the amount of render time allowed by 1/2

1 Like

You’re missing the point
@ForSerious surely meant that there likely were no updates to the model itself - just allowance of larger inputs.

2 Likes

It fixes the fringing in the ‘unplugged’ videos of Kurt Cobain, which look like VHS quality ones, but it doesn’t improve the details that much - the model tries to preserve the blurriness, probably. But it does this job better, than any single model I’ve tried on this video so far, it’s my fault to have the source in higher resolution, whereas it should be 480p, most likely. :slight_smile:

4 Likes

Yes, I see Starlight has done well reducing chromatic abberation. I feel like it will do well in reducing aliasing and moire.

1 Like

I think Starlight did quite well here.
This is the kind of upscaling I like - better a bit fewer details than those artificially pumped up images that scream “I’m AI” directly in your face.

It does keep the face and expression here very natural and still does a good overall cleanup of the material giving it some extra sharpness.

But then this video is REALLY FAR in quality from VHS (I guess you didn’t see a VHS in some time, otherwise you wouldn’t have rated this as VHS-quality) - in fact the source already is quite a bit higher than most DVD releases of concerts I know.

And, unfortunately, with such very bad, really VHS like material Starlight does heavily struggle in my first tests.

1 Like

This is one of my favorite examples of how Starlight can handle high-resolution but aliased and noisy footage.

3 Likes

Yeah I agree. I prefer a little less detail with some dehaloing to make it look natural. I’m still hoping for a local model that can get rid of aliasing and moire for line skipped video.

So this is another try with a real VHS rip:

Again, really terrible.


The after image looks like a different guy, the text (“5”) is terribly distorted - and look at the Hand / guitar neck :scream:
Oh, and why does Starlight desaturate the image?

So it seems that Starlight currently is for medium to higher quality SD and mainly HD material but not for really LQ sources.

P.S.: This is what IrisLQ can achieve:


Still altering the face but overall MUCH better than Starlight (text, hand, guitar - and even face).

5 Likes

In this case, it is likely caused by the colorspace handling in the MP4-only early access demo.

Full deployments of Starlight on cloud/server will include full codec controls

1 Like

Wonder what happens if you do Iris first and then send it through Starlight.

That’s an idea - might try that :+1:

1 Like

I’ve noticed this as well. My Starlight test clips need to have saturation increased by about 30% to match the originals.

1 Like

Just tried it here. Not a significant difference taking a 480p source to 1080p. Will be interested in seeing what happens when/if Starlight can eventually do a 4x upsample.

1 Like