What I wondering is, the main goal seems to be improve quality of Starlight, as the title says, as the CEO wrote. But for now I can’t see improvements since release. It’s not faster, or when it is, then quality is lowered. The Topaz Team said, they have to find out how to manipulate Starlight model, they can give us paramaters…but there is still no presets today we can do.
Instead improvements, they released Starlight Sharp, which is different but not better ( I think it’s worser than SLM), so where the roadmap goes to, more models instead improve the main, quantity than quality? I also can’t see efforts solving alien fonts issue and other deficits. I think Starlight is great but I just don’t see any progress here.
I tend to disagree with you on Srtarlight Sharp - I think it is better as usually the objects or hair are more defined which is great for low quality footage. One drawback of Starlight Sharp can be faces when compared to the normal Starlight but only when they are far distant or in bad input quality.
SLS is the new training model, and I think it’s slightly faster and higher quality than SLM. In my tests, I’ve seen SLS deliver better quality results. However, the real problem is that it’s still insufficient and extremely slow. It still distorts text and renders details artificially. So, a new training model is needed, and settings need to be added. There are still no settings for SLS and SLM. For example, VRAM remains idle during processing, indicating an optimization issue. It’s also possible to further improve performance by using a combination of CPU and GPU.
I can’t do post processing with SLS every try did worsening, while with SLM I can do, this is one reason I prefer SLM. Starlight sharp does invent more new things (textures) than SLM, this can be wanted or be a problem, pends on what you want.
So Starlight Sharp seems at first look better, because SLM looks (too) smooth, but this can be corrected. I do often dual-SLM (upscale 2x and again 2x), first upscale into 7.0.0.4b with 35GB GPU Ram usage (4GB Shared), then Iris MQ+Gaia (1x) as finisher and Scale+FineDehalo+add FilmGrain into Hybrid.
