Project Starlight - Topaz Labs
Posted: Sun Mar 02, 2025 12:32 pm
Topaz Labs released a major new model very recently. They claim that it "transforms low-resolution and degraded video into HD quality". Some descriptors from them:
Spielen wir Liebe: [Code] Mords Pas On t'Aime: [Code] Before/After comparison slider for Spielen wir Liebe: [Code] Before/After comparison slider for Mords Pas On t'Aime: [Code] Personal notes: Some dust/scratches were picked up on an removed, but it's clear that you would need a program specialised for that task to ensure complete/satisfactory removal. The enhancement is a bit aggressive for me, especially in terms of faces, but it does do a great job with the background; I think it would look great with some grain introduction from Topaz under the Gaia model (low invasive). Overall, promising results, I'm interested in your thoughts!
It's the first-ever diffusion AI model for video enhancement and the only one to achieve full temporal consistency, ensuring seamless motion from frame to frame. Users can access the research preview online or in Video AI.
Take almost any video to 4K: Project Starlight uses diffusion technology for the clearest AI video enhancement to date. It upscales, enhances, denoises, de-aliases, and sharpens video—without any sliders or manual adjustments.
Each frame restored by Project Starlight is created by analysing hundreds of frames surrounding it. This creates a smooth quality that has never been achieved in AI footage restoration before, called temporal consistency.
While existing video enhancement models utilize GAN technology, the switch to diffusion is one key to Project Starlight's major quality boost. Unlike existing GAN models, Project Starlight’s models have a great understanding of semantics, what objects actually are, and how motion and physics work—so they can make things look more natural with less to work with.
As stated above, it currently takes twenty minutes to process ten seconds of footage using server-grade hardware (apparently H100 GPUs, which are specifically designed for AI workloads and go for around 30K each). It's therefore only currently available as an online preview operating on a credit model (it would currently cost over $1,500 worth of credits to enhance a 1h 30m film), unless you happen to have a H100 GPU/cluster and are willing to pay an enterprise subscription, then they can allow you to process locally. They will be working on lighter models in the near future for consumer local processing:In 2023, we tried improving quality on our video models while holding speed and size constant, but this approach was ineffective. While later iterations (Rhea, RXL) do offer improvements over earlier versions (Proteus, Artemis), you’ve probably noticed that the changes have been getting more incremental. So in 2024, we asked the question: “How much video quality could we achieve if we didn’t care about speed or size?” The result is Project Starlight. It requires huge VRAM and currently takes 20 minutes to process 10 seconds of footage, which we know is sort of ridiculous. The results, however, are truly mind-blowing - an accomplishment which was quite challenging just by itself.
I've enhanced two short clips of moderately degraded quality with Starlight: One from Spielen wir Liebe (1978) and one from Mords Pas On t'Aime (1976)In 2025, we’ll focus on optimizing Starlight in a similar way. We’ve only focused on quality for this release, and we wanted to get it usable as soon as possible. You can expect smaller and less expensive models in the near future, based on the same technology that we’ve developed with Starlight. In the meantime, we’re initially pricing this model at cost to help make it more accessible.
Spielen wir Liebe: [Code] Mords Pas On t'Aime: [Code] Before/After comparison slider for Spielen wir Liebe: [Code] Before/After comparison slider for Mords Pas On t'Aime: [Code] Personal notes: Some dust/scratches were picked up on an removed, but it's clear that you would need a program specialised for that task to ensure complete/satisfactory removal. The enhancement is a bit aggressive for me, especially in terms of faces, but it does do a great job with the background; I think it would look great with some grain introduction from Topaz under the Gaia model (low invasive). Overall, promising results, I'm interested in your thoughts!