yes it’s not “mini” it’s maxi! 
RDNA 4 made some big changes. It goes toe to toe with a 5080 in VAI and closer to a 5090 then it should be…
yes it did made a big step up, in terms of smart power. but most of complex models still don’t support radeon. in general. most of the time - devs just don’t bother with it. since nvidia has all the ai market. Too much time being behind, several gen’s of being behind starting from rx 5700, that can’t be fixed with a couple of upper midrange cards. specially if we consider the fact that rdna 4 has no future, as the next one - udna - will be entierly diffirent. So there’s no point in optimisations for rdna 4, rdna 4 doesn’t even have the big top gpu. But i always say good luck to amd, as nvidia completely lost their minds with 50 series.
3 Likes
looks good. Now run RheaXL 1x in automode over the Starlight result 
But owning an Nvidia GPU was never a requirement to using VAI before. Nvidia cards are looking less and less appealing with each passing generation. The value proposition is awful. Meanwhile, AMD cards are quickly catching up. Why wouldn’t you build a cross-platform model like the others (if not now, then soon)?
2 Likes
I’m as nvidia owner totaly agree that their path for now is wrong, even if we not talking of money, prices, they depricated 32bit physx for no reason in 50 series. Thats all true. But the harsh truth is - nvidia has build the base for such a compute tasks decades ago, inventing cuda, the problems with amd and working loads are not new at all. It was like that for years. Any, and i mean any serius workload software was supporting only cuda, even 15 yeara ago, not opencl. The reasons for that are arguable, but it’s just a fact. And this state of things established long before nvidia became ngreedia triliondollarworth money-monster.. With neural networks it is just became more obvious for a lot more people. You’re sayng about cross platform models, but maybe you don’t know, most of those new fancy ai models are not cross platform. Go to any ai enthusiasts forum or reddit, you will see crying amd owners there. There is no such thing like cross platform model, there is cuda model, and directml model. building startlight for amd would mean a sereous rework, and i woudn’t count on it in short term. When they will optimise everything for cuda, implement all the settings, features, then i think they build some kind of directml implementation. i don’t think it will be done soon.
1 Like
I’m just a user and not deep into it, but the base is stable diffusion, or something else? AMD optimization is in progress
It’s so bad for TVAI users, we need tensor cores, are in the consumer segment, but the focus of NVidia here is on gamers playing just 1080p or 1440p. NVidia says “ok we save money and 16GB VRam is enough here” if you want 32Gb VRAM buy the 5090 for 3000$ 
5 Likes
I’ve seen examples of the Starlight Mini, and indeed it works well, provided you put up with a performance gap of more than ten times compared to older models.
I think enhancements to the picture have run their course, not just that it’s difficult to progress with existing effects, but more importantly that low definition video is becoming less and less common as equipment is upgraded.
The difference with this is that frame rate limitations remain strong, and as users seek higher resolutions, they often need to compromise on frame rate. The video rendering side of the equation is also still mass-producing a lot of low-frame-rate video, and more work needs to be done on frame interpolation and it will be more prevalent.
1 Like
bith would you like to work for me?
I need a global pause button for the render queue. When I render multiple videos and pause the current one, the next one starts automatically. This prevents me from actually stopping the rendering process to use other programs. A master pause button is needed to halt the entire queue, which can be resumed later. This is a very important and simple-to-implement feature.
Also, the “Starlight” process needs a pause button. It’s a very long process and we can’t take breaks. Please add these features in the next version.
5 Likes
@kevin.wang-topaz
Dear Kevin,
in the last patch release notes, your team said the performance for high-end systems/GPUs would be significantly improved for the Starlight Mini model. Pardon me if I disturb you and put this to your attention, but in my case, the performance stayed the same (1.0 fps at minimum setting and 0.4 fps at x3 upscale). I have a really high end PC and a rtx A6000 PRO Blackwell (the new flagship of nvidia, 96 GB VRAM). The render time stayed exactly the same as before last patch, however, the VRAM consumption dropped a bit, by 20% apprx which worries me, as more VRAM usually means more precision in the calculations. What did your team have in mind when your patch message quoted “high-end systems” if I may ask ? Thank you for reading, and kindest regards.
2 Likes
Less than 1.0 fps on an RTX 6000 PRO Blackwell… Holly Molly! Well, there it goes the need for anything beyond an RTX 4090 right now. Starlight Mini results are phenomenal, but the software needs improvement when it comes to processing speeds.
5 Likes
Yeah, I guess the team needs to tamper a bit with newest pytorch 2.7.1.+cu128 and have a look at sage attention wheels (the 2+ version will be published in a few days) to see how they can include that in workflows for high end Blackwell GPUs. That said, I am not angry, the team warned us the rendering times would be long with Starlight Mini. I just suspect they haven’t yet really looked into Blackwell architecture…
3 Likes
@tony.topazlabs
I have not yet found an up-to-date roadmap for TVAI in general and Starlight Mini (local) in particular at Topaz.
Is it foreseeable that Starlight mini will still be usable locally on an Apple Silicon Mac in 2025?
4 Likes
rest of 2025 roadmap coming?
1 Like
Hope we get Starlight Mini and Astra locally on M1 MAX. Don’t care if it’s 0.001 fps as my videos are only one minute in length. I’ll glady wait two weeks for rendering a one minute video.
2 Likes
I am in awe of what this thing can do. Not sure 0.001fps is good enough for anyone to be honest, you might say that now, but the electricity and heat will only shorten you hardware’s life. If my maths is right that’s 6 hours per second of footage, you will probably grow old before you see the result.
3 Likes
I’m on Apple M1 Max so the power consumption is A LOT lower compared to NVIDIA cards. Apple M1 Max is very power and energy efficient even under load, plus I use a cooling pad underneath for lengthy renders. Also, the vids I plan on doing are only one to two minutes max so don’t mind waiting a few weeks for it to render.
More power to you my man. I am on an i9 32GB with a 4090 and it’s still not fast enough, but this is by far the best up-scaler I have ever used.
1 Like
SLM + Astra if we get this model for local use, is definitly not practicable, doubling already horrible rendering time. We must find other ways to tune our SLM Results. Iris MQ or Rhea works but are invasive.
1 Like
