Personally, on my sequences, I don’t have any problems with vertical or horizontal lines or even tiling with StarLight Mini. And I do x2. If I do more than x2, it will really take too long for a few seconds of footage.
1 Like
Yes, until x2 I don’t have these vertical/horizontal lines in smooth surfaces or other artifacts.
They are only get “added” if I go above 2x.
Indeed, also with my RTX 5090 it takes too much render time to go above x2.
In the Topaz Video GUI - you can click “Process → Show Export Command” or “Process → Show Preview Command” and this will give you the ffmpeg command to run with the Topaz ffmpeg binary.
For example:
"tvai_up=model=iris-3:scale=0:w=720:h=480:preblur=0:noise=0:details=0:halo=0:blur=0:compression=0:estimate=8:blend=0.2:device=0:vram=1:instances=1"
The filter to apply is the tvai_up followed by the parameters for that filter.
Setting scale to 0 allows it to automatically scale based on the provided width and height.
The rest of the parameters depend on which ai model you select for enhancements. In the example command, iris-3 is the model being used. The preblur, noise, details,halo, blur, compression, estimate, and blend parameters map to the sliders shown in the GUI under “Enable Parameters” checkbox.
Again, to use the tvai_up filter, you MUST set the TVAI_MODEL_DATA_DIR and TVAI_MODEL_DIR as environment variables for the script, or use the command prompt window from “Process → Command Prompt”.
The available models for the model parameter correspond to models shown in the GUI.
But if you want the full listing look in the corresponding models folder -
On Windows: C:\ProgramData\Topaz Labs LLC\Topaz Video AI\models\
On OSX: /Applications/Topaz Video AI.app/Contents/Resources/models
The device, nvram, and instances all correspond to hardware encoding and GPU/hardware acceleration utilization. device=0 means auto - let ffmpeg decide automatically what to use.
Again - the important thing is to use the bundled ffmpeg to be able to use the tvai_up filter.
1 Like
Upon further investigation I found out that runner.exe gets its command-line parameters with a default output filename:
"C:\ProgramData\Topaz Labs LLC\Topaz Video AI\models\\runner.exe" D:/Temp/LEAR/King_Lir_akt2.mkv D:/Temp/LEAR/King_Lir_akt2_9_starlight_mini.mp4 --start-frame-idx 41573 --end-frame-idx 41751 --ffmpeg-encoding "-c:v h264_nvenc -profile:v high -pix_fmt yuv420p -g 30 -rc cbr -b:v 60M -preset p6 -movflags use_metadata_tags+write_colr" --ffmpeg-preproc-filters scale=w=768:h=576,setsar=1 --upscale-factor 2 --max-gpu-mem 8
Right now it appears the 5060 Ti is pretty optimal for SL mini. Easy to cool. 16GB of actually fast VRAM for about $500. Anything that would eat more than 16GB would be unrealistically slow to run on it anyways. Meanwhile from what I can tell the 4060 Ti struggles a bit due to the lower VRAM bandwidth, making the core itself wait more than optimal.
Ridiculous what they want for new graphics cards.
I shopped forever to find a 5070 ti…
Finally found one for $850….but even that is over MSRP.
What scares me even more is the possibility/excuse Tariffs will affect this ridiculous price more if they ever get fully implemented.
there are many issues with this type of input videos. you cant rotate and crop at the same time. I reported it but no reply and no fix. Its either ignorance, one man show or just the business model they are following. Unfinished unpolished UI replaced with a new one with more glitches and issues in order to justify new major version. Workaround is to use ffmpeg to prepare the source data (removing rotate metadata flag -metadata:s:v rotate=0, rotating pixels with lossless transformation -vf “transpose=1” -c:v prores_ks -profile:v 4). Most of the preprocessing inside of topaz Labs Video is just lame and not maintained.
3 Likes
so I asked perplexity (type of chatGPT LLM) and it couldn’t find any comprehensive official videos for topaz labs video 7. I guess that can tell us a lot about keeping up to date with support. I guess some videos are for v6 and v7 same though I would expect new features to be explained and have a use case comprehensive videos for them.
There is no official, comprehensive video tutorial series specifically for the new GUI of Topaz Labs Video AI 7 as of July 2025. Most available tutorials either cover version 6, general features, or individual updates, but not a full, step-by-step walkthrough of the entire new interface.
- YouTube Tutorials:
- Topaz Video AI Tutorials Tips and Tricks (Nov 2024) covers many workflow tips and features, but not a full series for version 71.
- Topaz Video AI - Upscale from 720p to 4K (May 2025) demonstrates upscaling in the new interface, showing the main controls and export process2.
- How to use AI Upscaling in Topaz Video AI version 6/7 - part 1 (May 2025) covers the basics of the new interface and workflow, with a focus on upscaling and model selection. This is the closest to a version 7 walkthrough, but is not a multi-part, exhaustive series3.
- Topaz Video AI Guide (Feb 2025) gives a general overview of the application, including the new tab workflow, sidebar controls, and codec settings, but is not a detailed, chaptered series4.
- Official Documentation:
- The Topaz Labs Video AI Quick Start Guide and Video Tutorials page provide written instructions and some video links, but the video series is mostly for v6, not v756.
Any update on whether these optimizations could be implemented? I would love to be getting 3-4fps for starlight mini.
- FP4/INT4 alone can deliver 4x to 5x throughput on supported hardware, since 4-bit ops run far faster and use less memory bandwidth.
- Structured sparsity could add a 1.5x to 2x gain, as half the weights are skipped in pruned layers.
- Transformer Engine v2 boosts efficiency via smarter precision tuning and op fusion, giving another 30–60% improvement on top.
These optimizations are multiplicative, not just additive, when properly implemented. So going from ~0.5 fps to 3–4 fps or more on a 1080p to 1080p video is plausible without degrading output quality, assuming quantization and pruning are done carefully.
7 Likes
Once these companies get a taste of cloud revenue they tend to de-prioritize local use in favor of it. Will be sad to see Topaz fail, but having customers pay for products twice (once under subscription then under a price gouge closed cloud use model) is a fast way to do it - especially when local use can be made 5 times faster on modern architecture.
The excuse about compute and model size is a poor one. They just need to allow instances to run in third party cloud and you can assign compute at 10X less cost and scale to crunch any size model.
It’s starting to look shakey but I’m still a fan of Topaz. I just hope the MBAs don’t kill another company I like.
1 Like
I’m not a shill for the company…heck I had several days of cursing it during my version 6 trials and tribulations.
Is Topaz dragging its feet to maximize and optimize SLM??
Dunno…but logic says….
Highly doubtful. Makes no sense to do so.
When they can’t even get their online main model “fast”….
And what they have given so far is pretty damn impressive. Finally an extremely simple model that does absolute wonders on older video.
Something so many of us have been asking for for a very long time.
Yes it is slow….very slow….the downside….and it only works decently on pretty high end Nvidia hardware….another downside….but results are really impressive.
Does Topaz see their online model as a tremendous source of income in the future….a product for the non techy, everyday Joe who just wants his old movies a heck of a lot better and can’t afford or even want expensive computer systems so he can to do it for himself?
Absolutely.
Can the two coexist successfully for the company?….
Sure.
What I see the biggest challenge for Topaz besides speeding up and lowering the cost of their online initiative ….is to rid themselves of their SLM total dependence on Nvidia.
Adding AMD and Apple to SLM I would think would be a MUCH higher priority for them then “optimizing SLM”…as it will increase their product base potential exponentially…and add the machines of Mac video Professionals….who whether PC enthusiasts want to admit or not….there is a majority In the video production world. But never fear…“Optimizing SLM” will just naturally happen as the technology matures.
All this gloom and doom and frustration talk….
Hell, I was ready to basically give up on Topaz a year ago with all the version 6 fiascos.
What I see coming from them right now….well it’s pretty darn marvelous!
8 Likes
From my point of view the Starlight online model is absolutely unrealistic from the very first start.
I have hundrets of videos made in SD quality (640x480 or 720x576). Each video has a total running time between 30 up to 60 minutes.
In addition I have thousands of SD video clips with in average 5 up to 10/15 seconds.
To upload all the SD footage step by step is one thing. I don’t know, what would be possible in the future to upload a huge number at same time as batch into the Starlight cloud.
Another thing are the costs (per minute). I don’t like to estimate the theoretically costs in advance 
But I think, they would call up some thousands of dollars in my case
Therefore I invested some weeks ago in a second PC with RTX5090 only for rendering Topaz models in the background.
But to be honest: I gained with the RTX 5090 not so much render speed.
But in comparison to my RTX 4070 it improved a lot, from 0,1-0,2 now to 0,9-1,0 fps using input SD and output “minimal” or “x2”.
I also like to have the Topaz work in my own hands in order to control the results during rendering. I don’t like to be “surprised” about the results coming back from cloud rendering.
2 Likes
Do you see any noticable (better) quality in the outcome with the 5090 (appearance, sharpness, color etc…)??
No, it is the same as with the RTX 4070.
It’s only “much” faster.
3 Likes
Sometimes the export is cancelled due to an export error. This happens in particular with larger files after more than 7 days of rendering time. The files created up to this point are immediately removed from the export opener by Topaz. Is it possible to make these files visible again and try to restore them? It is very annoying when more than a week of rendering time is wasted.
4 Likes
I went from a 4060 to a 5060ti to a 5070ti and now waiting for a 5090. I truly think it’s a disease although I’m stopping at the 5090 (I think) and will most likely keep it. 
4 Likes
Yup….5090s….a $3000 disease…
That Medicare, the VA, Tricare, Blue Cross-Blue Shield….and any health insurance company will NOT cover.
Darn it! 
4 Likes
Agreed. You’re talking about two different customers. People who prefer to run things in the cloud and people who prefer to run locally. I will NEVER use cloud, no matter how fast it is. I prefer local. I hate cloud for any kind of processing. I hate one drive and all that trash. If Topaz moved to a cloud only service I would simply unsubscribe and stop using their products. As Hulk Hogan once said, “That doesn’t work for me, brother.”
9 Likes
It is precisely because 5090 is so expensive that I am more willing to use a 4090 local cloud server with a monthly rent of $126, so that I do not have to invest such funds all at once.
1 Like
I never thought, that I would spend on one day so much money in a graphic card
But on the end, I’m very happy that I invested on a second PC with a much more faster graphic card for all “matters” in video editing.
As you said: I’m not willing to repeat such high investments in a graphic card again and again
2 Likes
