Starlight on AppleSilicon is broken even more in this “release” (??) than before. Not only delivering abysmal quality like the 1.0.3 now there is an additional bug RGB shift giving quite “psychedelic” results.
So, this is supposed to be a release version, right? It would be bad enough to have such in a Beta but you must not, never ever have something like that (totally broken functionality) in a release version!
Is there ANY testing done on Apple Silicon at all?
Sadly, all of Topaz Video’s Starlight optimisations are being directed towards Nvidia GPUs, so Apple and AMD platforms are being treated as an afterthought. I need to face the reality that I’ll need to upgrade to a high-end RTX card, probably a 5080.
Hmm, I don’t know.
I still do like the good “old” TVAI 7.0.2 the most when doing SL mini upscales on my NVIDIA equipped PC and those newer versions aren’t better, neither quality-wise nor regarding speed.
So, in conclusion I didn’t see any worthy improvement in SL at all, sometimes more downgrades.
At least not as bad as on the Mac where they ruined a slow but working model to a state where it’s totally useless.
That sounds a to me a lot like the Windows 11 SSD issue that’s been happening in some systems.
I have two nesr-identical systems (both Ryzen 9, same RAM capacity and even the same motherboard chipset), but the Nvidia system is stable, but the AMD one isn’t, and that’s only since the latest update. I think Starlight is totally broken on AMD but perfectly fine on Nvidia.
That’s totally possible. I haven’t tried the beta on it yet, but I got a 9070 XT, and this 1.0.4 version won’t even download Starlight.
That’s because the AMD version of Starlight needs 20GB of VRAM, so it won’t work on the 9070. It did work on my 7900 XTX, but not anymore.
Got it. Everyone’s reporting that it gives worse looking results on AMD. Since I have already tried SL on all my videos that the other models still don’t improve meaningfully, and didn’t like the results, I don’t really feel like I’m missing anything.
yeah, borked here too, and i can’t get previous versions to work anymore. i updated to macos 26.1, is that what broke it?
Probably not going to get many likes on this, but updates have clearly gotten worse since they added support for AMD and Apple systems. Now their focus is split 3 ways, instead of focusing on optimizing the more popular Nvidia cards. Updates used to be slow as it is when it was just Nvidia. Now they have to support two different additional architectures. I’m looking at a report that says Nvidia has a 92% market share in 2025. Resources are being split to support cards that have a much smaller market share and it’s now affecting the Nvidia optimization and performance.
Topaz is clearly having issues optimizing all 3 different architectures, as seen in the last few patches that haven’t fixed the big issues. One issue gets fixed, but three more are created.
I’m not saying Topaz should drop support for those cards, because they marketed the program to work with those cards and people have paid, but it is what it is. If you care about running ai workloads, you really should be using an Nvidia card.
It was never just NVidia but also AppleSilicon and AMD all the time from the start.
Only very recently with the Starlight model it was NVIDIA but also only at the start. And even there people claim that later versions were less capable than succeeding ones (so development wasn’t really there even though there weren’t AMD/Apple versions).
And, for your “market share”: if you’re looking only at the gamer’s market this is maybe true but not for TopazVideo: just look at the dl counts for this version: 61 Mac downloads vs. 162 windows ones. And from the windows count you’d of course have to subtract the AMD share. So Apple users still do make a VERY considerable amount of Topaz’s User base.
Also, most of those NVIDIA users don’t have the top of the line 409/5090 cards but lesser ones that perform similar to the faster AMD and AppleSilicon rigs.
When Topaz Video Enhance AI (VEAI) launched in February, 2020, it only supported NVIDIA GPUs on Windows. If you had an AMD GPU, you were limited to CPU-only processing, which was painfully slow and impractical for most users.
You can see my own old post from back then, where I was pleading with Topaz to add AMD GPU support:
AMD GPU support was finally added starting with version 1.7.0, released in November 2020. Here’s the official release notes thread, which clearly mentions the new AMD compatibility:
However, it didn’t take long for me to realize that relying on an AMD GPU for AI-specific tasks was a shortsighted choice. The vast majority of AI software—especially open-source projects on GitHub—relies on CUDA, leaving AMD users with workarounds or dead ends.
NVIDIA’s dominance extends beyond hardware. The company frequently dispatches engineers to partners and projects at no cost, helping them optimize software for its GPUs. In stark contrast, AMD often ignores outreach efforts. Ultimately, this relentless ecosystem-building has cemented CUDA as the de facto industry standard, leaving AMD far behind.
yes, and that’s exactly what’s sad - no improvement since release, rather a downgrade. I find it particularly tragic when Topaz makes a new release where minor bugs are fixed but the new destroys other things that worked before, so then better do nothing 
Unfortunately this is the only way to work with any AI now.
All the models are trained at NVIDIA systems.
Moreover, the models are trained at 80GB Nvidia H100 at least and require a bit of trickery to work with measly 24GB of VRAM.
Even 5090 with 32GB of VRAM is not enough to generate 1080p videos without tiling.
Remember guys when Topaz promised Intel Arc compatibility? Pepperidge Farm remembers.
I’m currently investigating multiple upgrade options, including an M3 Ultra Mac Studio which includes 96GB of unified RAM on the base model. Unfortunately, I’m probably looking at spending close to TWICE the price of an RTX 5090! May just pick up a cheap 5070 Ti and call it quits!
That’s why I use 7.0.0.4.b for SLM, only there I have 100% Vram usage + 5GB shared (doing 960p) and the model draws 39GB gpu ram going greater than 960p.
Maybe it’s just my imagination that the quality is better, but I think it is.
unfortunately i still haven’t got starlight to run in the cli, can anyone help please? I want to go from sd to 1080p and not 960p, so scale 2.25
Any news about new starlight version with even better quality than starlight?
Don’t remind them. 
