Uploaded December 2021 | Updated November 2026, 4 days ago
NOTE: This version uploaded to YouTube using HEVC Rec. 2100 PQ. (10 bits) 1000 nits maximum.
Since most of my latest Unreal Engine project have already been converted over to 4.27.1 or 4.27.2, I was bummed to find out that unsurprisingly since UE 5 Early Access was release in July, 2021 with the last update coming in early September that levels created or modified in 4.27 would not show up. To alleviate that I went out to Unreal's Github and downloaded the latest source build (12/7) and created a new UE 5 that could open it. However, it also seems Lumen had some terrible glitches handling lights and particle occlusion that caused it flash-leak out of occluding box. So I went back and creating another build based a later date (12/16) and now I had a whole host of new problems since emissive textures were 10-100 times brighter, and sequencer file kept freezing, requiring me to kill the process. Soooo, I just created new sequencer, sequence, which I wanted to do anyway for a new "Hot Cocoa in VR" trailer video based on some new "friends" to keep you company in the long winter months.
In working with Lumen, I find it very easy but frustrating since even though everything is real-time, many of my material had to be tweaked to behave. It should also be noted that I wanted to do this without any NVIDIA RTX features so I remained on DX-11 so what you see is ALL Lumen and few Nanite meshes.
To render this out, I used both the legacy 'Movie Scene Capture' and 'Render Queue'. The former was used ONLY for rendering audio since sadly 'Movie Scene Capture' still does allow for submixes. The video was rendered as an EXR (Linear 16 bits per color) image sequence with 3:2 Temporal Super Resolution anti-aliasing. This image sequence was brought into Adobe Premiere using the EXR Pro plugin and and set the project up for (203 58% PQ), then adjusted white and black levels to the HDR PQ levels of 1000 nits. For export I chose HEVC at 40 mps Rec. 2100 PQ
Even though this was more of a study in using Lumen & HDR, most of it useless in VR since Lumen & Nanites do not work there. My next test is to see if they omni-capture tool can, that way I can capture 7K images using specially designed capture cinema camera.
You can download the file used for upload to YouTube to view directly on your HDR monitor: https://1drv.ms/v/s!AjCOOLquXxqngvcdjzwXSieV2DeriA?e=oI6THn
NOTE: This version uploaded to YouTube using HEVC Rec. 2100 PQ. (10 bits) 1000 nits maximum.
Since most of my latest Unreal Engine project have already been converted over to 4.27.1 or 4.27.2, I was bummed to find out that unsurprisingly since UE 5 Early Access was release in July, 2021 with the last update coming in early September that levels created or modified in 4.27 would not show up. To alleviate that I went out to Unreal's Github and downloaded the latest source build (12/7) and created a new UE 5 that could open it. However, it also seems Lumen had some terrible glitches handling lights and particle occlusion that caused it flash-leak out of occluding box. So I went back and creating another build based a later date (12/16) and now I had a whole host of new problems since emissive textures were 10-100 times brighter, and sequencer file kept freezing, requiring me to kill the process. Soooo, I just created new sequencer, sequence, which I wanted to do anyway for a new "Hot Cocoa in VR" trailer video based on some new "friends" to keep you company in the long winter months.
In working with Lumen, I find it very easy but frustrating since even though everything is real-time, many of my material had to be tweaked to behave. It should also be noted that I wanted to do this without any NVIDIA RTX features so I remained on DX-11 so what you see is ALL Lumen and few Nanite meshes.
To render this out, I used both the legacy 'Movie Scene Capture' and 'Render Queue'. The former was used ONLY for rendering audio since sadly 'Movie Scene Capture' still does allow for submixes. The video was rendered as an EXR (Linear 16 bits per color) image sequence with 3:2 Temporal Super Resolution anti-aliasing. This image sequence was brought into Adobe Premiere using the EXR Pro plugin and and set the project up for (203 58% PQ), then adjusted white and black levels to the HDR PQ levels of 1000 nits. For export I chose HEVC at 40 mps Rec. 2100 PQ
Even though this was more of a study in using Lumen & HDR, most of it useless in VR since Lumen & Nanites do not work there. My next test is to see if they omni-capture tool can, that way I can capture 7K images using specially designed capture cinema camera.
You can download the file used for upload to YouTube to view directly on your HDR monitor: https://1drv.ms/v/s!AjCOOLquXxqngvcdjzwXSieV2DeriA?e=oI6THn



![IMPRESSIVE RESULTS on CachyOS/SteamOS under $1200 w. Frameworks Strix Halo miniITX board w. DLSS 5
By now you probably read about the several mods for bringing modified DLSS 5 Neural Engine Models to AMD Radeon RX90x0 and RX70x0 for Windows, but did you know you can do the same with gaming Linux build like CachyOS, Bazzite and even SteamOS?
Over the past couple of weeks I have been testing a Linux specific AMD Neural Engine mod for RDNA 4 and RDNA 3 (works with 3.5 as well). Not only was I surprised on my under $1000 Frameworks Strix Halo - Ryzen™ AI Max 385 - 32GB which is a slightly less powerful version of flagship AI Max 395, but with 32GB unified memory at 8000Mhz, and a slightly faster clock speed so you wont notice the difference with gaming, especially due to its dynamic memory management giving you between 8-22GB of VRAM allocation. Since its powered by the 8050 GPU (a few less cores) it is only slightly less performing than the 8060 GPU in the AI Max 395 but for WHOLE LOT LESS in price. This is housed in upcoming housing/fan/power/Oculink solution Gaming Lens: SFFMU Jewel Series that is smaller than a Valve Steam Machine in cool translucent jewel colors with RGBA custom lighting to make it glow (if you want).
With that I have been able to run most AAA games in the past two years with AMD RDNA 4.1.2 INT 8 at 4K between 30fps-70fps with varying super scaling settings and graphics settings at high to ultra. Since it runs over 60-90fps at 1920x1200, I thought it might be interesting to see how well it runs the latest DLSS 5 Neural Engine adaptions for AMD RDNA 3, AND WAS PLEASANTLY SURPRISED. 40-60 fps with modified experiment performance models and as simple as installing Optiscaler to Win64 game executable folder (noted exception to Phantom Blade, which I talk about later). I have now got to work with Silent Hill Townfall, Resident Evil: Requiem (with ReShade), A Plague Tale: Resonance, and now Phantom Blade Zero.
Is it worth it? For me who is more of casual gamer (at 67) with no real twitch skills, these games are more for the narrative and immersion which for me the more photo realistic the better, and frankly I dont need more than 30 fps as long as the latency isnt too bad. Every game I tried it with I like a lot better with the DLSS 5 Neural Engine mod than without it. Of course I dont over do the effect and use the character mask and ensure I have good game depth map, but even a light touch is great. When I was viewing cut scenes from Silent Hill Townfall, I felt like I was watching a movie. Especially the female character Zoe Ellis. Worth video upload just for that scene.
Anyway, here is link to the mod I used by saaaaaaaaaaaaaaaaaas
https://github.com/saaaaaaaaaaaaaaaaaas/DLSSNR-RDNA3-0.0.8.3/releases
Just follow the instructions and should go pretty smoothly, but be mindful that Phantom Blade Zero needs it installed in the Win64 game folder with changes to the Optiscaler.ini settings to change paths back from the default set by Optiscaler that points to another folder with shim executable.
NOTE: To get the DLSS 5 Neural Engine settings to come up, press your END key. This Optiscaler mod is based on plugin with its own Ui. Sadly in my case the cursor went away, so you are flying a little blind, but keyboard directional keys will help navigate the menu. The setting changes are in real-time, so you will see immediate results. You can even change neural engine models on the fly [F11] to see which one looks or perform best and even turn it off [F6] and do comparison screen shots [F5].
Also I use DLSS-Updater (https://github.com/Recol/DLSS-Updater) for Linux which allows the latest NVIDIA, AMD and Intel graphics .dlls to be cached and updates ALL or piecemeal .dlls, as well as updating the NVIDIA Neural model if it finds it. Do this before starting the installation Optiscaler Neural mod and make sure you find the lastest nvngx_dlssnr.dll (the script asks for it). This way Optiscaler will automatically choose FSR 4.1.2 INT 8 for RDNA 3 and 3.5.
Use your own gaming folder paths based on the Proton/Wine Windows path configuration for Phantom Blade Zero
[Libraries]
;
; Main folder for OptiScaler to check dll files below
; Default is .OptiScaler
OptiDllPath=S:steamappscommonPBZProjectPBZBinariesWin64
; Options below for overriding path per dll
;
; nvngx.dll or _nvngx.dll
NvngxPath=auto
; nvngx_dlss.dll
NvngxDlssPath=auto
; nvapi64.dll or fakenvapi.dll
NvapiPath=auto
; amd_fidelityfx_dx12.dll or amd_fidelityfx_loader_dx12.dll
FfxDx12Path=auto IMPRESSIVE RESULTS on CachyOS/SteamOS under $1200 w. Frameworks Strix Halo miniITX board w. DLSS 5](https://i.ytimg.com/vi/UrCwrXGuoeI/mqdefault.jpg)






