118
you are viewing a single comment's thread
view the rest of the comments
view the rest of the comments
this post was submitted on 08 Oct 2026
118 points (99.2% liked)
Steam Hardware
23343 readers
187 users here now
A place to discuss and support all Steam Hardware, including Steam Deck, Steam Machine, Steam Frame, and SteamOS in general.
As Lemmy doesn't have flairs yet, you can use these prefixes to indicate what type of post you have made, eg:
[Flair] My post title
The following is a list of suggested flairs:
[Deck] - Steam Deck related.
[Controller] - Steam Controller related.
[Machine] - Steam Machine related.
[Frame] - Steam Frame related.
[Discussion] - General discussion.
[Help] - A request for help or support.
[News] - News about the deck.
[PSA] - Sharing important information.
[Game] - News / info about a game on the deck.
[Update] - An update to a previous post.
[Meta] - Discussion about this community.
If your post is only relevant to one hardware device (Deck/Machine/Frame/etc) please specify which one as part of the title or by using a device flair.
These are not enforced, but they are encouraged.
Rules:
- Follow the rules of Sopuli
- Posts must be related to Steam Hardware or Steam OS in an obvious way.
- No piracy, there are other communities for that.
- Discussion of emulators are allowed, but no discussion on how to illegally acquire ROMs.
- This is a place of civil discussion, no trolling.
- Have fun.
founded 5 years ago
MODERATORS
No, this is not possible. FSR 1 is an algorithm that works on the output image, like a Photoshop filter. That means HUD elements can't be distinguished from the actual game world in example. Since FSR 2 (and upwards) its different. It needs to be integrated into the game, understanding what it does and compare multiple images. It is Ai based and requires training from the game itself. That's why you don't get system level integration for DLSS too. So that is out of question.
Afaik dlss and fsr are pretrained models and they don't train them for each specific game. Modders totally do enable them for games that don't support them natively, but it can be quite involved because it needs the motion vectors and maybe depth from the game.
Maybe I misunderstood it and its not about adjustment training to specific games, and its about adjusting settings or something like that. I searched for a resource to point to that, but cannot find. It's really hard to search about this topic online, because its always the same basic explanations what Ai upsaclers are (probably written by an Ai model itself...).
I know of modders being able to replace an existing model that already has motion vectors and depth. It would just reuse those information, because the API is agnostic to the models. If modders add vectors and depth by their own, that does not change the fact that DLSS or FSR 2+ can only work with those data.
I think it needs depth buffer access and a UI mask, but not necessarily deep integration beyond that? At least for the FSR3 shader code I’ve looked at personally.
This is not out of the question for generic integration.
I’m less sure if FSR4 though, as I think that was a pretty big leap over 3.
I don't know the exact process too, so we are both speculating. But its fact that Temporal upscaling (that is FSR2 and upwards for FSR related ones) need integration from the game. And these are Ai models that needs some kind of training data from the game as well. I don't know what that exactly means, but my guess is that the integrated Ai model to upscale is trained with the data in the game and then delivered as a package. That makes it impossible to integrate a general purpose Temporal upscaler with Ai into a system, that could be utlized on any game (or program if you will). So I don't know how this cannot be anything else "out of question". I mean this is by concept not possible, as far as I understand (and why nobody did it too).
But an Ai system could be used as a Spatial upscaler, like a filter that is not integrated into the game. FSR 1 itself is not using any Ai, but DLSS 5 has such capabilities (think of the new lighting model, it could be tweaked to upscale the image instead).
I don’t think it’s actually trained per game. Maybe per type of game.
I do remember Nvidia offering individual game finetuning as a service to game devs for a hot minute, but really, these models generalize well. A model they spent time tweaking and training on a big basket of games might work better than a finetune for one specific game, as it would just get overfit and perform worse in edge cases.
I’ve seen this happen in older GAN models, and newer diffusion ones, and especially in LLMs. Even with the exact data the model is going to work on, finetuning can make it perform worse with the actual stuff it’s going to miss or mess up.
I can use FSR4 in games that don't have it, on hardware that doesn't even support it with Optiscaler. How did it train on shit it wasn't actually made for if that's how it worked?
What exactly do you mean? Any link? As explained. FSR 2, 3 and 4 needs training data from the game. It is possible to swap DLSS out with FSR variants, because DLSS has the training data already. But you cannot have FSR 2, 3 and 4 on all games like FSR 1. And FSR 4 also has to modes as it seems, one classical Temporal like 2 and 3 that runs on relatively older hardware too, and the new ML-based variant that requires the new cores on RX 9000 hardware.
If you use Optiscaler, then I think what happens is that you swap the DLSS implementation with FSR 4 Temporal variant out. I searched the web a bit and found this: https://github.com/optiscaler/OptiScaler
So what part I was wrong about exactly?
You swap it with whatever upscalers it already has, yes. But not all games have any FSR or even DLSS support. Some only have XeSS, and you can replace that with FSR or DLSS using the utility. If it never had any kind of FSR support what-so-ever, how is it able to work at all?
I did not say this, where do you get that? I explained that FSR 1 is different from all other upsalers, it is not comparable to all of the other methods. FSR 1 is like a Photoshop filter, while the other methods have builtin data from game. You cannot add a Temporal Upscaler to a game that does not have any. You cannot use OptiScaler on a game that does not have any upscaling at all, it needs DLSS 2 or higher, or FSR 2 or higher, or XeSS, because those have the training data using Ai with the game builtin. You cannot use OptiScaler without any of these, or with FSR 1.
What exactly am I wrong about?
I am not saying you're wrong, I'm trying to understand what you posted. But I see what you mean now. I was under the impression the training data was specifically for the specific upscaler so FSR4 would need at least FSR2 supported natively.
Oh, no no, I did not say or imply that. I was concerned about FSR 1 and the difference to the other upscalers. And the difference between the second new mode of FSR 4 that is not available on any other. So even if you have DLSS or FSR 3, you can only use the Temporal variant of FSR 4, not its ML-based variant (if I understand it correctly). At least that is what I said, so its probably true (self irony).