The First-Ever “Local Desktop” Diffusion Model For Video Enhancement

is the second diffusion-based upscaling model Topaz Labs has released, but is the 

. Starlight Mini brings what makes the Original Starlight (now 

) so special, into a more efficient variation for local processing. 

Like it’s predecessor, it is capable of enhancing footage that may otherwise be a challenge with our more traditional AI Models. This can be anything from tricky 8/16mm film to de-interlaced mini-DV footage.

This AI Model is definitely the most powerful enhancement on the desktop application. Based on this criteria, it is reasonable to expect hours of rendering pending your hardware.

Not everyone will have access to the minimum requirement or time required for this AI model on their desktop machine, so access to this AI Model is also optionally available for cloud processing on server-grade graphics hardware.

So what can you expect out of Starlight Mini when compared to the the Precise Model? The Precise model simply has a more sharp and "clean" quality to it's outputs due to the fact that it still requires such immense resources and hardware in order to provide outputs. The desktop variant, Starlight Mini, is essentially the optimized version where you get the essentials out of Precise, but you gain in it's uniqueness that can be an even more desirable output. This equates to some slight lower finesse when it comes to details, however on the whole, the results appears more natural and consistent.

Don't get fooled by the "Mini" part of the name. The name refers to the more efficient make-up of the model that can run on the local desktop. Quality of this model has been considered better than the Precise Model by several users!

To get the best results out of this model, it's good to know what it's intended for and what makes it unique.

For older videos, results are at generally better than the Rhea model with more comparative details, subtle definitions and consistency in the animated motion. It is without doubt the best Desktop AI model for older footage enhancement/upscales.

A unique look that comes from this model, is a subtle haze, which is great for human faces and more natural looking footage. This is similar to having a diffusion gel in front of a harsh light on set. This greatly reduces the “plastic” look of upscales and enhancements that are too clean.

Additionally, this model also greatly reduces flickering commonly seen in flat backgrounds that cause banding.

We recommend testing standard definition (SD) videos that have less than ideal results with our previous models, though it’s not limited to that resolution. This will give a larger “Wow” moment. The difference between standard definition vs high definition will naturally be more drastic. Essentially starting with something very good in quality may have marginal results.

Still frames of the original video clip 316x240 (left) and the Starlight Mini output 1280x972 (right).

Starlight Mini had dropped the “Mini” after the original Project Starlight was renamed in ASTRA (our web app) as Precision Quality. Precision Quality, could only render on the cloud due to the complexity and size of the model.

Starlight “Mini” was originally named for it’s more efficient model size that still keeps a high quality output.

With Project Starlight renamed, it was thought best to promote the Starlight Mini model’s name to just “Starlight” for the desktop experience. This model produces outputs that rivals the original Project Starlight while adding some unique properties that is more natural to the original content. So, the “mini” part of the name was removed.

Unfortunately, this only caused confusion. We’ve listened and learned. It gives us great pleasure to return the “Mini” back to this model’s name. By a large request, the Starlight Mini name is back to stay!

, represents a series of AI models. Models such as Starlight Mini, Starlight Sharp, etc are variations of this series with specified properties to enhance and improve your footage.

Here’s a quick start guide on how to use the Starlight Mini model.

To begin, have your clip already imported.

 to ensure that the graphics card in use is 

 (if you have multiple cards) and be sure to select the appropriate codec for your shot. *AMD and MAC users can now use a variation of this model.

Starlight Mini is the first AI Model on the list.

If this is your first time using Starlight Mini, a pop up menu will appear (NVIDIA only). This will run a system check on your machine to see if it is capable of running Starlight locally. Click on the “Run System Check” button.

This will give you a visual check mark for every step that’s required. Otherwise you’ll get an “x” that identifies if there is an issue with the machine to prevent the large download for this Model.

Once all the necessary checks are approved, you can then press the Download button. 

Once completed you can close out of this menu.

Assuming that you’ve already installed Starlight Mini, when this model is selected, you will be presented with the following notification:

 use this AI model without any other enhancements nor filters. Your machine will be working very hard with Starlight alone. 

Check the button to reimport the video. Once the render is complete, the result is essentially used as a new source. Here you can control additional enhancements, rescales and filters.

Starlight local shows different options for resolution output from the model, based on the resolution of the input video.

Once a resolution is selected, either export locally with 

To ensure the best quality outputs, Starlight exports will have a minimum resolution of 1280x720.

If your footage output settings are under 1280 width, then the full resolution will move up to match the 1280 width.

If your footage output settings are under 720 height, then the full resolution will move up to match the minimum 720 height.

Aspect ratio of the video content will be kept the same. Note, at this time, the Letterbox/Pillarbox will be applied.

Recommended to use square-pixel video footage at or under 1920x1080.

From the desktop, Starligh Mini/Cloud rendering supports AV1, VP9, H264, H265 video codecs.

The web input, however, only accepts H264 MP4 files.

If required, you can always re-import your Starlight Mini result to scale up/down your footage or remove the Letterbox/Pillarbox.

Auto Crop to remove the Letterbox/Pillarbox here

While this is new for all and we continue exploring and refining the Starlight Mini model, there are a few things to consider when using this tool:

You will need at least a Pro account to render locally. 

A requirement of an NVIDIA GPU with at least 10GB of video memory, but we strongly recommend a GPU with 16GB or more memory for optimal performance and quality.

One of the minimum NVIDIA GPU requirements, is that the GPU needs  the compute capability of at least 7.0, which means all the legacy GPUs with compute capability smaller than this version is not supported, including GTX 10 series and Quadro P series.

In addition to these GPU requirements, Starlight also requires a minimum of 32GB of system RAM.

It is recommended for AMD, to have 20GB of VRAM.

The recommendation for Mac is to have a minimum 36GB of RAM.

This is an AI model that is both complex and resource heavy. Meaning, that it could take hours to process.

It is not recommended to run Preview or Frame Preview with this model. For Nvidia users it will not work. AMD and Mac users this may also not work. It is recommended to export a few seconds for Preview purposes. 

Per Cloud Render, there is a 5 minute maximum for Starlight.

Sound is not supported on the Cloud at this time.

Please note that if you have text that is too small, blurry or inconsistent, the result may not be "sp3ll3d" correctly. If you can barely make it out, so will the AI model. It may assume that the text is a texture or pattern.

Water may appear smooth when compared to source.

If you have interlaced videos, it is recommended that you deinterlace first before enhancing with Starlight.

Have a great time using Starlight Mini. Enjoy the unique qualities from this “first of it’s kind” diffusion model for the consumer desktop!

Starlight Mini

Starlight Sharp excels at processing low-resolution footage, delivering superior results for challenging content. The model provides improved handling of smaller faces and facial details for enhanced clarity, while producing noticeably sharper and more detailed enhancement results compared to previous offerings.

The following available outputs are multiplied relative to the original resolution or minimum/maximum resolution threshold for the model:

*Starlight Sharp, can export a 4K output using a special second pass. This is using a Nyx variant under the hood to assist in the 4K output. Alternatively, You can decide your own 4K upscaler by inputting the 1080 result as source.

Generates a more realistic render with an approach to minimize the “rubber-like” output found on human faces when other upscalers or restoration AI Models are too ‘clean”.

Reduces some compression flickering from low quality videos.

Starlight Sharp can Produce a bit more details when compared to Starlight.

: Available on both cloud and local platforms for maximum flexibility

: Leverages multiple NVIDIA GPUs in parallel for enhanced processing capabilities

Just a little below, you will see a section called 

This is where you will click on the caret to reveal the drop down menu. The second option is 

If this is your first time using the Starlight Model, you will be prompted to “Run system check” before you can download the model.

Once you click on the button, the system check will conclude if your system can indeed run the model and you will be given access to download.

When the download is complete, you can close out of this window.

Now that you have the model downloaded, you can now use it on your videos.

Starlight Sharp is currently only available for local rendering to Windows users with NVIDIA GPUs. 

This model is quite demanding in terms of VRAM, requiring a bare minimum of 16GB or more in order to run the model effectively.

It is also recommended to have 32GB of RAM.

Due to the model complexity, this is a large resource intensive AI Model, so long render times are expected when rendering locally.

Recommended to use square-pixel video footage at or under 1920x1080 to avoid any aspect ratio shifts

Ensure that you are using the latest driver for your graphics card.

Currently capped at 1080p resolution in the cloud. 

Starlight Sharp is a standalone enhancement and can’t be combined with other filters.

It is not recommended to run Preview or Frame Preview with this model. It will not work. It is recommended to export a few seconds for Preview purposes. 

Please note that if you have text that is too small, blurry or inconsistent, the result may not be processed correctly. The AI model may interpret unclear text as a texture or pattern.

If you have interlaced videos, it is recommended that you de-interlace first before enhancing with with this model.

Starlight Sharp

Introducing Starlight Fast 2, for local desktop rendering! Starlight Fast 2 is the third variant to the desktop Starlight Model Family. Note that Starlight Fast (formally Starlight Speed) exists on Astra. This model is great at getting good diffusion results at a much faster processing time. Note, that in order to make this model both available locally and still be speedy (for a diffusion model) there is a slight difference in quality between local and the cloud iteration.

If your content is already at the max output, such as 4k, then no matter the scale factor, the max output will be rendered:

Best results come from 2x and above upscaling.

Great option for faster local diffusion based rendering.

Recommended to use video inputs at or under 4K.

This model works for Medium to High Quality content whether this is generative or live action videos.

Speed can be 3-5x faster than Starlight Mini pending on your GPU and available VRAM.

Provides better treatment for flickering/stuttering frames (especially when camera moves)

Model access via the cloud for all subscribers.

This model is less ideal for content with:

Haloing issues common with vintage digital videos

hair quality can vary in detail when content is on the lower medium spectrum.

Not intended for very low/poor condition quality media.

Interlaced content needs to be de-interlaced first.

Just a little below, you will see a section called 

This is where you will click on the caret to reveal the drop down menu. The third option is 

You will need a Pro account to render locally.

Starlight Fast 2 is currently only available for local rendering to Windows users with NVIDIA GPUs.

There is a slight quality difference between renders with 16GB vs 32GB VRAM. The latter being a tad more detailed.

Though this model has been noted to be 3-5x faster than starlight mini, it is still a diffusion model, so long render times are expected when rendering locally.

Recommended to use square-pixel video footage at or under 4K UHD to avoid any aspect ratio shifts

Currently capped at 4K resolution locally and in the cloud.

Starlight Fast 2 is a standalone enhancement and can’t be combined with other filters.

It is not recommended to run Preview or Frame Preview with this model. It will not work. It is recommended to export a few seconds for Preview purposes.

Enjoy faster diffusion rendering with Starlight Fast 2!

Starlight Fast 2 (Local)

, for local desktop rendering! Starlight HQ is the fourth variant to the desktop Starlight Model Family. This diffusion model has a great balance of detail and sharpness while keeping the intent of the original video source. Provides higher quality outputs for a wide range of videos, though not necessarily intended for archival content it can still provide good results. Though addressing similar intentions as Starlight Fast2, outputs have generally been higher quality and consistent with this model’s longer processing time.

Processing speed is very similar to Starlight Mini.

Requires a minimum of 10GB VRAM though 12GB VRAM and up is recommended.

Model access via the cloud for all subscription models.

Haloing issues common with vintage digital videos.

Hair quality can vary in detail when content is on the lower medium spectrum.

This is where you will click on the caret to reveal the drop down menu. The fourth option is 

Starlight HQ is currently only available for local rendering to Windows users with NVIDIA GPUs.

This model is quite demanding in terms of VRAM, requiring a bare minimum of 10GB or more in order to run the model effectively.

It is also recommended to have 12GB of RAM.

This is still a diffusion model, so long render times are expected when rendering locally.

Starlight HQ is a standalone enhancement and can’t be combined with other filters.

Starlight HQ

 made its debut in our Cloud app, Astra, with much excitement and positivity. In the meantime, we were also working hard to get this wonderful model for local desktop rendering! I’m happy to announce that the team has just done that! The best part, we’re making this model available for everyone that meets the requirements to run it! This makes Starlight Precise 2.5 the 5th variant to the desktop Starlight Model Family. Note, that previous iterations of this model exists on Astra. This model has been specialized for realistic human face and skin textures in addition to other areas of improvement such as material, texture, and text.

This model works for Medium to High Quality content whether this is generative, live action or used for restoration videos.

Improves generated content that may have soft, plastic, or slight artificial looks, however the AI generated content should have solid structure and composition for ideal improvements.

There is a slight difference between the local version and the cloud version in order to fit on the desktop.

Greatly reduces flickering with smaller textures.

Cloud Export is supported on both Mac and Windows.

Expect long render times for this diffusion model.

Available for all licenses to run locally! Yes, you read this right!

Requires an absolute minimum of 12GB VRAM though 16-24GB VRAM is recommended.

Targeted for GPU compute compatibility of 8.9. This is also the recommendation. Please note, though there might be some success with a CC of 8.6 cards, this is not technically supported and some may encounter a less ideal experience.

This model is included locally, using Neuroserver.

Requires a minimum of 10 frames for ideal outputs, though 15 frames is recommended.

Shots with heavy motion blur may be rendered as texture.

Starlight Precise 2.5 is a model variant of 

Starlight Precise 2.5 is currently only available for local rendering to Windows users with NVIDIA GPUs.

This model is quite demanding in terms of VRAM, requiring a bare minimum of 12GB or more in order to run the model effectively.

It is recommended to have 16-24GB of RAM.

Recommended for GPU compute compatibility of 8.9.

Starlight Precise 2.5 is a standalone enhancement and can’t be combined with other filters.

 to run Preview or Frame Preview with this model. It will not work. Export out a few seconds for Preview purposes.

Though this model does handle text and logos much better than previous Starlight Models, please note that if you have text that is too small, blurry or inconsistent, the result may not be processed correctly. The AI model may interpret unclear text as a texture or pattern.

Starlight Precise 2.5

Plugins

Advanced Functions

Understanding the different Starlight AI models that Topaz Labs has to offer within the Topaz Video application. Be informed to create the best possible output for your projects.

All diffusion models (Starlight Family) require Neuroserver: Starlight Mini, Sharp, Fast 2, HQ and Precise 2.5.

Neuroserver installation happens through the installer with a checkbox. An in-app update is recommended.

If for some reason the Neuroserver is not working as expected, reinstall the app to run the “Repair” mode.

Where are all the current Starlight Series of models located?

 (desktop app) houses all the Starlight AI models that can be rendered locally.

 (cloud app) houses the Starlight AI models that are cloud rendering only.

Learn more about the original Starlight AI Model that started it all! 

Starlight Series