RTX 5060ti 16gb Optimal settings for 360p → 4k
SLM 3x 360p → 1080p
SLP 1x 1080p
Proteuse Natural 1080p → 4k
Davinci Resolve Add Film Grain Using Dehancer Pro As Desired
2 Likes
Yes, I agree with you.
I think the Optimal workflow For 360p is SLMx3, then run it through SLP x1 and then Proteus Natural pass if you need 4k.
I’ve been gathering the workflows people have been posting here and I organized them. I haven’t been able to test them yet. Can we figure out if one of these is objectively the best?
Workflow 1 (360p to 4k)
SLM 360p → 720p
SLP 1x 720p
NYX XL 720p → 1080p
SLP 2x 1080p → 2160p
Workflow 2 (360p to 4k)
SLP 3x 360p → 1080p
NYX XL 1x 1080p
Proteus Natural 1080p → 2160p
Workflow 3 (360p → 4k)
SLM 3x 360p → 1080p
SLP 1x 1080p
Proteus Natural 1080p → 2160p
2 Likes
Seedvr2 7B models are not good. Try 3BQ8, much sharper and more “creative”. Also the higher the upscaling, the less jittery it is, so a 4x upscale is smoother and creates more details than 3x, which is better than 2x. 1x enhance with SEEDVR2 is so jittery it is unusable. I think you really need 32gb of vram to use SEEDVR2 properly as I heard it was originally developed for 96gb of ram. With a RTX5090 you can really just manage to do 1080p properly with a max batch size of 77 and temporal overlap of 6 frames at batch boundaries to avoid vram oom errors. If you have a RTX 6000 pro with 96gb, you can do 1440p with probably batch size of 200 and the results will be much smoother and rendering time will be much quicker due to larger batch size. Topaz should consider allowing manual setting of batch size so pro users with larger vram can optimize temporal consistency and reduce jittery motion - this ensures RTX 6000 or RTX 5090 are not stuck running at the same quality and speed of small batch sizes that were tuned to work on RTX 5070 or RTX 5060.
7 Likes
Yeah I’m sure the quality would be much higher with more VRAM allocation.
I only have a 3080 12gb, with a 4:3 video I achieved 1600p (from 480p) with a batch of 17 frames with 6 temporal overlap frames, it looked pretty good and I used the other models such as 3BQ8 though text seemed to become muddy using that model.
But yeah SLP adds a bit more quality per batch frame than SEEDVR2 models did for me personally.
Currently out of money to even think of upgrading any part in my system atm, (considering the price hike recently).
Maybe in the future with the 60 series I’ll upgrade and hopefully get 32gb VRAM. 
Hot take:
There’s no such thing as the one perfect workflow—and honestly, there probably never will be. Source material varies way too much, and everyone has different preferences anyway.
Hardware matters a lot too, especially when it comes to render times. Even with my setup (Ryzen 9950X3D, RTX 5090, 64 GB RAM), there’s no way I’m running something like SLM → SLP → Proteus Natural → SLP. The time cost just isn’t worth it. And from my experience, stacking a bunch of models doesn’t automatically give you a better result.
I also don’t really get the whole “dirty faces” complaint about SLP. Yeah, it adds more noise than SLM—but it also keeps a lot of the fine detail that SLM tends to wipe out with that waxy, over-smoothed look.
Personally, I don’t like that look at all. It feels artificial and kind of lifeless. Noise and grain are part of what makes an image feel real—they add depth and texture, especially when you’re heavily upscaling.
Sure, SLP can hallucinate some skin detail, but I’ll take that any day over a plastic-looking image with no edges or imperfections.
My current workflow—even for something like 240p—is pretty simple: SLP at 4× for anything up to 480p, then 2×–3× for 480p, and 2× is already enough for 720p.
After that, I run Proteus Natural with some grain and maybe a bit of noise to hit the final resolution (1080p or 2160p).
5 Likes
Yes, contrary to theory the 3b models deliver higher quality than the larger 7b ones - and stay away from FP8.
Still starlight precise beats SeedVR hands down.
4 Likes
hands down. i try starlight from every single angle. before this neuroserve, SEEDVR2 it was a great option to start, and to enhance, but by far topaz give us a banger. I hope that they now, focus only in SLP and keep update the model to like, 10FPS.
3 Likes
Latest beta was in march, any new betas? Can’t wait to try something new
1 Like
For the foreseeable future, the emphasis in development should be getting the VRAM requirement down. Until the AI bubble bursts and businesses stop trying to replace every human being on their payrolls with AI tech, GPUs and VRAM aren’t going to be getting more affordable. Nvidia is reportedly even bringing back some of its 30 series GPUs, that’s how bad things are in the supply chain.
3 Likes
I’m curious why anyone thinks SLM as a first pass is a good idea. SLP is going to do a better job bringing out the small fine details in text etc.
3 Likes
What are your grain/noise settings in PNat
This looks wild, Topaz Team. Been doing Topaz stuff since 2009 and along for the sometimes rocky ride with TVAI. Cannot WAIT to try it out!
1 Like
VRAM utilization and speed improvements would be awesome.
SLM first pass only for under 480p footage without Text
1 Like
I agree. To me it seems much more logical to run SLP first on the original source file, and then SLM for the finalizing 4K upscale, as SLP digs up the most of the original “hidden” detail. Especially for low rez/quality sources.
For this to function optimally, though, to avoid the 25 frames visible updates of SLP, I am now running a music video upscale test using hi8upscale’s special 5-export workflow to see how well this works…
And grain in which case? SLP gives me WAY more actual details in this very low-res video (384x192 pixels) that I am running tests on right now… Much more than SLM does based on the original source file.
Ah now is this an interesting way to do it. I like.
Depends a lot on the source I guess.
In my case the video is grainy, low-rez and shot handheld (but with little compression).
The handheld factor plays an important role in SLP digging up loads of sub-pixel details…
Doing SLM on it as the first and only step worked out okayish, but the background details were so-so. SLP as the first step does it waaaay better on it.
I might show you the differences when the new processing is done.
2 Likes
