首先~這不是我寫出來的，也不只是能用於 Anima，它可作用於所有模型。什至不是AI出的圖也能用，突別適合用於3D及2.5D的舊CG. 一般的 Realistic Enhancer LORAS可能會改變圖片的結構什至把部份如手部等變壞而且也費時，這是一個高速的 Realistic Enhancer 方案。
First of all, this wasn't written by me, and it's not just for Anima; it works with all models. It even works with images not created by AI, and it's especially suitable for older 3D and 2.5D CG. Typical Realistic Enhancers like LORAS can alter image structure, even ruining parts like hands, and are also time-consuming. This is a high-speed Realistic Enhancer solution.
作者是"叁贰玖零",codon2001
The author is "叁贰玖零(3290)".codon2001
=======================================================
這是一個用一秒就可增加圖片真實感的方案
This is a solution that can enhance the realism of an image in just one second.
文件節點:https://github.com/codon2001/comfyui-dlssnr-test
Custom node: https://github.com/codon2001/comfyui-dlssnr-test
目前只支援 RTX 40 及 50 系列顯示卡
Currently, only RTX 40 and 50 series graphics cards are supported.
下載文件只是個工作流，請自行在上方的Github連結安裝及貼上對應的"nvngx_dlssnr.dll"
Downloading file is just a workflow . Please install it yourself from the Github link above and place the corresponding "nvngx_dlssnr.dll" file
安裝方法:
How to install:
use below command in CMD
cd /path/to/ComfyUI/custom_nodes
git clone https://github.com/codon2001/comfyui-dlssnr-test.git
cd comfyui-dlssnr-test
..\..\..\python_embeded\python.exe -m pip install -r requirements.txt
============================================
你可以在網上很容易就找到"nvngx_dlssnr.dll"，但記得它分為RTX50 及 RTX 40兩個版本，請下載對應你的顯卡示的版本。
You can easily find "nvngx_dlssnr.dll" online, but remember that it comes in two versions: RTX50 and RTX40. Please download the version corresponding to your graphics card.
Description
Comments (20)
@ppkkmoon Can You provide on/off examples :) ?Is it worth hassle?
updated ^^
How does it impact gen times?
The change is subtle in the examples, is there a slider to increase it?
@N0n4m3 As shown in the examples, maxing out the sliders produces only minor variations in 2D and photographic styles, but triggers dramatic changes in 2.5D and 3D CG renders. Designed as a 1-second high-speed realistic enhancer powered by DLSS 5, its main purpose is to refine details rather than generate new images from scratch
@N0n4m3 This isn't just another standard realism LoRA.
@N0n4m3 https://civitai.red/images/141376166 <- I think it is worth to give more 1sec to give the improvement, but U may not want it..... ^^
@ppkkmoon Thank You for the answers.
Mainly I was interested in seeing if DLLS5 could be used to actually generate whole scene but it appears that they cut that functionality down so it can run, barely, on single 5090.
I think it would be nice if DLSS5 could work on anything, like as in any app, be it video / image / game / ai gen, like some sort of a filter. I would like to see how Quake would look like, but there is already RTX for that. That being said, I never used Nv Ansel so... ^.-
@N0n4m3 I believe we should have reasonable expectations for DLSS5; it's reasonable to assume it can only improve 2.5D or 3D CG. Because DLSS5 is simply an AI filter, it can't significantly improve the realism of Kera2/Flux images. Nor should it drastically alter 2D images, as this would severely distort italicized images like those from Forza. Its strength lies in its high-speed realization of 2.5D or 3D CG, making it very suitable for Anima models. While Kera2/Flux images are realistic, NSFW conversion isn't easy enough, and Anima, despite having native NSFW, struggles to generate realistic images. DLSS5 is a perfect fit for Anima/Illustrious models...
@N0n4m3 I add a result with 3 step DLSS5 with 1 facedetailer => https://civitai.red/posts/30757291
As shown in the examples, maxing out the sliders produces only minor variations in 2D and photographic styles, but triggers dramatic changes in 2.5D and 3D CG renders. Designed as a 1-second high-speed realistic enhancer powered by DLSS 5, its main purpose is to refine details rather than generate new images from scratch.
I believe we should have reasonable expectations for DLSS5; it's reasonable to assume it can only improve 2.5D or 3D CG. Because DLSS5 is simply an AI filter, it can't significantly improve the realism of Kera2/Flux images. Nor should it drastically alter 2D images, as this would severely distort italicized images like those from Forza. Its strength lies in its high-speed realization of 2.5D or 3D CG, making it very suitable for Anima models. While Kera2/Flux images are realistic, NSFW conversion isn't easy enough, and Anima, despite having native NSFW, struggles to generate realistic images. DLSS5 is a perfect fit for Anima/Illustrious models...
I always get error:
- Exception Message: RuntimeError: Create DLSSNR Feature 18 failed (NGX=0xbad00001)
Even with correct DLL. I asked some LLM for help but they can't figure out why.
Anyone else got this issue?
Do not use my WF, try to use the exsample wf in the node foler. Because the aurthor had updated the code ^^.
@ppkkmoon Actually the LLM discovered the issue : the GPU selector lists the gpu based on their index in Torch, but the rest of the code uses the DXGI index for gpu instead. Basically I have two gpu, only one of them can run DLSS-5, and the code is bugged :)
@eicosa Hope U could have fun with it now ^^
@eicosa I'm sorry, there's definitely a bug — I used Codex to help write the code, I'll try to fix it.
@codon000 Hi~ 3290
@codon000 oh the solution is already here for you https://github.com/codon2001/comfyui-dlssnr-test/issues/2
@eicosa Really appreciate your help!
@codon000 and i appreciate your work! I literally went to bed the other night thinking "i wonder if possible to load DLSS 5 into comfy UI?" ... and when i woke up, it was done :O