Like I have an RTX 5090 and 9800X3D, but this thing still runs like the washing machine from the 19th century:
A. Starlight Precise keeps “loading model” FOREVER
I don’t have a screenshot for it but it happened when I tried to resume a task when the very same task HAD BEEN running for 1 and a half day and was cut off for NO GOOD REASON but a complete death to my PC: no screen, no keyboard, no mice and NO AMOUNT OF POWER BUTTON PUSHING could shut it down (never seen this in my entire like). I had to cut off the power and re-do everything.
B. Who is going to believe this is only a 1080p to 4k on a RTX 5090:
I knew Topaz Video AI from way back like 5 years ago (yeah it was called AI Enhance or something) and now I wanted to come back for its new features. Only disappointment is now. So did the team run those Starlight Precise 2.5 benchmarks on RTX 8090???
Logs show you are running into two known issues at this time that the devs are working on.
First, MKV containers are experiencing issues when being processed by the Starlight Precise model.
Second, using In/Out markers to trim a video for processing can cause it to load and stall without moving into the actual processing steps.
These are being worked on by the devs for a fix, and the recommendation is to either roll back to 1.4.0 or remux the file into MOV or MP4 containers. Then trim the file outside of Topaz Video and bring in the shorter clip, or process the entire file with Starlight Precise 2.5.
Topaz changed the underlying architecture of the Starlight models in release 1.4. If you were around in Topaz Video Enhance AI days, when they switched to using FFmpeg and the AI processing as filters, you might remember a tough transition period. We are going through something similar now.
The last fully stable release was Topaz Video 1.3. Install that build and Starlight Mini, Sharp, Fast and HQ will all work correctly. You won’t have access to Starlight Precise, but Sharp, followed by a quick run through Proteus will get you quite close. Hopefully we will get a release 1.7 or 2.0 soon that will resolve the problems with the new architecture.
Render times for all of the Starlight models (which are based upon a different kind of AI analysis than any of the other models) is most impacted by the size (resolution) of each frame, going in and out. For example, a 2X upscale of a 640x480 to 1280x960 will be faster than a 3X upscale to 1920x1440, both of which will be faster than 720x540 to 2880x2160.
Overall time is obviously also influenced by the frame rate of the original (i.e. a 24fps original will complete sooner than a 29.97fps source of the same runtime since there are fewer frames).
The average processing time for a 2X upscale from 720x540 to 1440x1080 at 29.97fps is about 41 hours on my PC (Intel i9 14900K CPU, RTX 5080, 64GB DDR5 system memory) using Starlight Sharp. I haven’t been able to run Precise on my main PC (same problem you have) but on an older machine (Intel i9 9900K CPU, RTX 5070Ti, 32GB DDR4 system memory) it runs about the same speed as Sharp does on that system. Typically, depending on the content being processed, I have seen 5090 users report anything from 0.5 to 0.9 fps. So far, I haven’t seen an actual report of reaching 1 fps on consumer GPUs. Perhaps a Blackwell RTX Pro 5000 (48GB VRAM) or RTX Pro 6000 (96GB VRAM) will get there, but all I have seen is conjecture since those cards are about $5,000 and $10,000 respectively.
Yes, in the final analysis, throwing more CUDA cores at the problem is the only thing that helps speed. More VRAM doesn’t make it noticeably faster, it just makes the output look better in the same amount of time (due to larger batches, and room for more overlap). But even there, the newer releases (1.4 and up) are not even using all of the available VRAM, basically capping everything at 5070 or 3080Ti quality level and higher spec GPUs offering only more CUDA cores. That is all the workstation cards offer - more cores than most of the consumer cards.
After I wrote in the forum over a year ago that I could see visibly better quality with SLM in TVAI 6.0.2b or 6.0.4b because my VRAM usage was maxed out, people replied that this was supposedly not the case. So I became unsure whether what I was seeing was real or just my imagination. According to them, and based on the known behavior of SeedVR, more VRAM only reduces tiling, which provides a speed boost but does not improve quality. So now I’m unsure what’s actually true.
The funny thing is, there wasn’t even a speed boost, because both of the above-mentioned TVAI betas not only maxed out VRAM usage, but also started using large amounts of slower shared system RAM. The highest amount I ever saw was around 49GB of total GPU memory usage in TVAI 6.0.2b on my 5090.
Anyway, that’s in the past now, but my logic says it’s never a good thing when software does not make optimal use of the available resources. I have a 5090, and today just the half of my 32GB VRAM is being used.
You could almost start thinking, in a cynical way, that Topaz is doing this intentionally so nobody gets a really fast system or significantly better quality with high end hardware — otherwise the cloud service would become less attractive.
As quality IS better with SLP on the Mac for local rendering with it’s massive amount of GPU RAM I believe it’s just that they settled for the most common denominator with those Nvidia GPUs.
Actually, I think they are sort of caught between a rock and a hard place. On the one hand, the cloud service may appear more lucrative, but is it really? The amount of hardware dedicated to cloud processing can only increase (if they are successful) which is expensive to bring on line, particularly now. Also, big jobs will tie up those cloud resources further applying pressure on resources. Bottom line, providing the online service is a double edged sword. I would submit that it is really designed for the casual user that wants to enhance some family video clips or the like and doesn’t have any computer that could possibly run the software successfully.
I can’t see any professional using cloud rendering - they already have the processing power locally and uploading their content opens up lots of privacy and security issues. The locally processed option is their only option, so Topaz will never go cloud only (at least not for the foreseeable future). Those of us that are not pros but use Topaz Video on our own computers for our own purposes basically ride along with the professionals.
You can see this in their release process. Professionals will never change software version in the middle of a project (at least not without a VERY good reason), so any new feature release may have bugs but they don’t really cause their main customers serious problems. By the time the pros are ready to implement a new release, many of the bugs will be worked out, discovered by the amateur users.
In fact, I would not be surprised if the Pro level tools start to use all available VRAM while the “basic” release is limited to 12 to 16 GB VRAM so that it works on as broad a selection of GPUs as possible. I think all that would be required is shipping a few different sets of BLOBs for various maximum memory configurations.
