LS 2.10.2 - spatial upscaling comparison

[ChaoLaChao]

Hello everyone!

I am proud to finally post a hefty comparison of the spatial scaling options available in the latest version of Lossless Scaling.

General:

For testing, I used my personal laptop: Lenovo ThinkPad E14 G1. It's a very GPU and bandwidth limited device, which allowed me to pin-point frametime and framerate differences between available scaling modes. The OS used in comparison was a fresh Windows 11 install (build 22631.3958) with the latest GPU drivers installed (101.2127). Game Mode was activated via Xbox Game Bar.

The app I did my tests on, on the other hand, was the official FSR 1.0 tech demo that AMD provides for everyone interested; I toggled native borderless fullscreen mode, enabled TAA and injected iMMERSE MXAO via Reshade 6.2 (all settings were set to their default values except for "Increase radius with distance" toggled on).

When scaling, I used the "Resize before rescaling" mode. Every scaling mode had its default sharpness level engaged (1/4 for LS1, 5/10 for everything else) and was using its respective "Ultra Quality" mode (which meant running the demo in 831p).

Specs:

CPU: Intel Core i5-10210U
RAM: 16GB DDR4, SODIMM 2667 MT/s
GPU: Intel UHD Graphics (TechPowerUp's GPU Database says it's an OC-d UHD 620)
Display: 1080p IPS, 60hz

Scaling modes:

• FSR 1.0
  
• FSR 1.0 Lite
  
• LS1 1.2
  
• LS1 1.2 Performance
  
• NVIDIA Image Scaling
  
  

  The comparison is available right below:

  https://imgsli.com/Mjg0NzA1/0/3
  
  

  tl;dr

• Native 1080p TAA = 15 fps
  
  
• FSR 1.0 = 21 fps
  It's solid. Its scaling pass has that trademark minor edge smoothing, while its sharpening pass is fairly weak and non-intrusive.
  
  
• FSR 1.0 Lite = 22 fps
  Great for iGPUs. Its scaling is less precise, while its sharpening pass has a tendency to leave haloing around the edges.
  
  
• LS1 1.2 = 20 fps
  The best one quality-wise. Its scaling has little-to-no edge smoothing (fine-tuned custom Catmull?), while its sharpening pass saves lots of smaller details.
  
  
• LS1 1.2 Performance = 22 fps
  The best one for iGPUs. Its scaling and sharpening passes are virtually the same as the ones seen in LS1, except for minor artifacting around less-defined objects (take a closer look at the pillar on the left).
  
  
• NVIDIA Image Scaling = 19 fps
  Lanczos scaling and DEAR LORD look at that AWFUL sharpening pass!!! Awful. Don't use it.
  
  

  Extras (FSR1/LS1 model detail comparison)

  https://imgsli.com/Mjg0NzE4

[Raigavin]

So FSR1.0 is better than LS1 for Image Quality?

[Spook]

Originally posted by Raigavin:

So FSR1.0 is better than LS1 for Image Quality?

IMO FSR's biggest advantage is that it doesn't tend to exaggerate sharpening artifacts. LS1 is better at keeping fine details fine. If a game/video suffers from oversharpening or overly-aggressive film grain, use FSR. Otherwise LS1 is technically better.

On the linked imgsli comparisons you can CTRL+Mousewheel UP to zoom in on specific parts of the image. The details on the gun demonstrate the differences most clearly.

[DDENN]

how do anime4k's setting work ?

[Zef]

Originally posted by Raigavin:

So FSR1.0 is better than LS1 for Image Quality?

Nope, LS1 is still the best and Nvidia's is just way way to oversharpened compared to the original image. FSR is just ok in most cases.

In 90% of usecases, LS1 is the best.

[kripcision]

Originally posted by Raigavin:

So FSR1.0 is better than LS1 for Image Quality?

personally i found LS1 to be too sharp, but my monitor may be exaggerating it since it has a crisp and sharp image mainly for fps games (so you can see the outline of people easier). so i found FSR to be better for me personally, but LS1 has its fair share of users that prefer it.

[Spook]

Originally posted by DDENN:

how do anime4k's setting work ?

Larger is better but increasingly more performance intensive. Turn on draw fps and/or your GPU-monitoring software and see what your GPU will take. Probably turn on VRS, mouse-over the toggle to learn what it does.

[ChaoLaChao]

Originally posted by Raigavin:

So FSR1.0 is better than LS1 for Image Quality?

Nope. LS1 is more worth it, unless you're playing a CPU-limited game (that's Source engine games from my experience lol)

Originally posted by DDENN:

how do anime4k's setting work ?

The bigger the model the higher the quality and resource usage. Just stick with anything below medium on iGPUs and you're good to go, although I don't see it worth it for anything sub-1080p.

[Xavvy]

Green teams NIS is pretty terrible but it's important to note that it's only the sharpening that's terrible, the up-scaling is fairly decent. The real strength lies in driver side full OS upscaling with 3% sharpness(I never use NIS in Lossless). But it's really only worth using at 4K or higher resolution. but at 4K+ plus it's pretty impressive at squeezing a bunch more performance out of a GPU with little to no image quality loss. I use 67% NVCP NIS at 3% sharpness at the OS level(enabling Image Upscaling(NIS) and then selecting the the 67% resolution forcing all of windows at this resolution. Now when you game use Quality DLSS combined and you basically get near Quality DLSS on native for the price of nearly Ultra Performance(which is really huge at 4K+ resolution)

Also using NIS driver side means you get no additional input latency since it's before the games entry in the pipeline. I've combined all driver side NIS(Quality)/in game DLSS Quality and LS1(Ultra Quality(1.3x)) and had pretty impressive performance while barely losing much for visual fidelity on a 4-6K resolution HDR1000+/Dolby Vision TV.

To be honest in my opinion spatial upscalers are best used in the 4K department or used at Ultra Quality settings in 2K range when the game has no FSR2+/XeSS/DLSS. They need a high base resolution to look good(At least 1080p imo as a base). 1600x900 doesn't look bad upscaled to 1080p on a smaller screen but imo that's the limit for me. I can't stand the loss below that. But imo at 2k base upscaling to 4-6k you really don't notice much of a loss because you're starting with such a high base.

Also values of 3-13% sharpening are okay in NIS NVCP as long as your resolution is high enough. Personally I use 0-3%.

[Spook]

Originally posted by Xavvy:

^^^

Yep, driver-level NIS is way more useable. I think LS's NIS implementation would benefit from more granularity in the shapness slider and the ignore film-grain parameter that you can set in NVCP.

[Xavvy]

Originally posted by Spook:

Originally posted by Xavvy:

^^^

Yep, driver-level NIS is way more useable. I think LS's NIS implementation would benefit from more granularity in the shapness slider and the ignore film-grain parameter that you can set in NVCP.

Yeah Lossless NIS is so different from NVCP NiS and yeah the sharpness control in Lossless is just not enough to combat the stupid sharpness method it uses. This is why I use 0-3% in NVCP because to be honest the Lanczos sinc that NIS uses is basically the pre version that DLSS is built on(it's a very good upscaling method)..

Obviously DLSS uses AI and tensor cores to leverage its magic but NIS is very good and comes with pretty adequate sharpness at 0% sharpness. like NIS really doesn't need any sharpness added at all in my experience and generally speaking you're better off adding sharpness via nvidia app overlay since it uses a multitude of better sharpness methods.