The model seems to only run in X3 mode.
Good to see the starlight mini released for local processing and try it on my set, 14700k, 4090, 64G RAM, similar to above, I got stable 0.5f/s from an 480p anime clip with 4.5 min for upscale x4, it took 6.5h, just went through the output and compared with the cloud starlight, on this particular clip mini is very close to cloud version yet it washed out a bit details, but I would say it was much better result from using other combination models (gaia+proteus., etc) with much clearer edge and lines, less AI artifacts and mystery lines on clear area, I am putting another 5min clip and wait for extra 7 hour to see 
.
Thanks @tony.topazlabs for explaining the option “Reimport export as source”, the current and ongoing test clip do not have this option checked, I can test combination model from starlight mini+other model to achieve best closer result to cloud version.
2 Likes
I prefer to save the first pass safely out of the workflow and run subsequent passes as new sessions. It’s why I don’t use the second enhancement option.
it does help to speed up from bulk processing when my first build become my second setup in garage 
The idea is that you install the graphics card in the second PC when you buy a new one. The second is slower, but it can run non-stop while you have the main computer free. You can do remote desktop access to the second. It always annoys me when my computer is blocked.
Yes, exactly, you got it.
I did my first Starlight Mini Test now and gets 0.3 fps bottleneck is my “470 none TI” with only 12GB VRam. As others mentioned before, the result is good, maybe on the level of his big brother. What I see is Starlight has still problems with fonts, here an example
the original
Starlight Mini does fixed 960p, but we need 1080p and 720p
2 Likes
Does starlight mini not have a preview mode? Do you just have to start an export? How do you use this model? I clicked on Enhancement then I chose my resolution, but I don’t see a preview button.
1 Like
Does it work on RTX 5000?
I get this error with 5090:
2025-04-30 19-26-30.101 Thread: 16020 Info FF Process Output: 9 C:\ProgramData\Topaz Labs LLC\Topaz Video AI BETA\models\Lib\site-packages\torch\cuda_init_.py:235: UserWarning:
NVIDIA GeForce RTX 5090 with CUDA capability sm_120 is not compatible with the current PyTorch installation.
The current PyTorch install supports CUDA capabilities sm_50 sm_60 sm_61 sm_70 sm_75 sm_80 sm_86 sm_90.
Starlight mini is not working for me either on a 5090. For pytorch, right now, I believe you need to use the nightly build. I’ve done this for other AI programs.
NVIDIA GeForce RTX 5090 with CUDA capability sm_120 is not compatible with the current PyTorch installation.
The current PyTorch install supports CUDA capabilities sm_50 sm_60 sm_61 sm_70 sm_75 sm_80 sm_86 sm_90.
If you want to use the NVIDIA GeForce RTX 5090 GPU with PyTorch, please check the instructions at Start Locally | PyTorch
To get comfyui working I needed to use torch==2.7.0.dev20250311+cu128 torchvision==0.22.0.dev20250312+cu128 torchaudio==2.6.0.dev20250312+cu128
1 Like
No there is no preview mode in this beta
1 Like
I’ve been testing the Starlight Mini model across various videos with different resolutions and quality levels. Overall, the results are quite impressive, but I’d like to share a few observations and suggestions based on my experience:
1. Visual Quality
- Eyes on low-quality footage: In some cases, especially with poor source material, the eyes can appear unnatural or even unsettling. This becomes more noticeable because the rest of the face is rendered quite cleanly, creating a strong contrast.
- Skin texture: The skin often looks overly smooth or waxy. Introducing a subtle noise or grain filter could help add more realism and improve the natural look of the output.
2. Performance
- On my system (7800X3D, RTX 4090, 64 GB RAM), performance ranges from 0.6 to 0.1 frames per second, depending on the input. Currently, the rendering time feels disproportionate to the result. I understand it’s still in beta, but I believe there’s significant potential for optimization.
3. Feature Requests
- Preview Option: A short preview feature (e.g., 1–2 seconds) would be extremely helpful for reviewing specific scenes before committing to a full render.
- Resolution Flexibility: The current minimum output resolution of 1280x960 feels limiting. I’d appreciate the option to go down to 720p, which would improve both performance and compatibility in many workflows.
4 Likes
I’m not having the same resolution limitations you guys are reporting.
I’m able to output to 720p, and it does indeed process faster. 0.7fps.
Maybe it’s because I’m outputting to png?
from 960p to 1080p is a small step upwards and just slightly slower, I think they should gives us both 1080p and 720p
2 Likes
did png now set 960x720 and get again 1280x960
That’s weird.
It is, I try different codec and input now
tests done, yep I get always the same 960p, tryed png,h264,265,ProRes and FFV1, different sources
I’m unable to get Starlight Mini to work at all with my 3090. It seems like a similar issue to what’s been previously reported—the runner process spins up, takes up all the VRAM, and then just crashes. But unlike the other reports, this is the error I get just prior to it crashing out:
2025-04-30 18-52-51.469 Thread: 53520 Warning qrc:/workflows/TProcessor.qml:442: Error: Cannot assign std::nullptr_t to int
2025-04-30 18-52-52.255 Thread: 53520 Info process exited: 4 1 0
2025-04-30 18-52-52.255 Thread: 53520 Info Going to delete path: "C:/Users/Andrew/Desktop/<filename>.mp4"
2025-04-30 18-52-52.255 Thread: 53520 Info Close and delete path "C:/Users/Andrew/Desktop/<filename>.mp4"
2025-04-30 18-52-52.255 Thread: 53520 Info Going to delete path: "C:/Users/Andrew/Desktop/<filename>_starlight_mini.mp4"
2025-04-30 18-52-52.255 Thread: 53520 Info Close and delete path "C:/Users/Andrew/Desktop/<filename>_starlight_mini.mp4"
2025-04-30 18-52-52.257 Thread: 53520 Info ~TProcess(): destroyed
2025-04-30 18-52-52.259 Thread: 53520 Warning qrc:/videomanager/TVideoManagerPane.qml:1058:29: QML IconImage: Detected anchors on an item that is managed by a layout. This is undefined behavior; use Layout.alignment instead.
2025-04-30 18-53-00.176 Thread: 53520 Info Logging files: 5
logsForSupport.zip (23.4 KB)
If the folder name also contains multibyte characters, starlight_mini will cause an error.
