Jobs Support Register

TW3 Mass Re-Texture with ESRGAN - Artificial Intelligence Texture Upscaling - Tool Needed!

For those unaware, this awesome AI neural erm... code/thingy (?) can upscale textures like nothing ever before with great results, but some results end up a bit weird. It probably won't replace real artists' work, but I can see it being used to mod older games, like TW3! Check out pumped shotties here -



If someone can write a GUI Tool for this - , then perhaps TW3 can become the first fully re-textured game via ESRGAN . It would be really cool. All those secondary textures not reworked by Halk Hogan could use some of that AI magic. I bet the end-result would be much better than my generic upscaling technique...
mmm .. to me it seems that does nothing but increase the sharpness of the image :p

it's a basic photo editing i think

EDIT: done in 2 minutes


Last edited:
mmm .. to me it seems that does nothing but increase the sharpness of the image :p

it's a basic photo editing i think

EDIT: done in 2 minutes
this reaches far beyond a simple sharpening in photoshop.
While some sample shots might look like it at a quick first glance, its really, really not.
Just look at the second pic again (the most dramatic difference imo). cant seriously say thats just sharpening lol.

To the topic in general:
i think this can achieve the most impressive results when the base material is really low rez (much older games), which defenitely isnt the case with witcher 3.
Still should be worth a try. Ill read through it, tho it appears an nvidia gpu is best for this (which i dont have)
I don't know about how well it works on textures with high resolution, but resolution and detail aren't the same thing. Some high resolution and even high detail textures look awful because of the mesh detail or just mesh area being too large.
so, just to report back @MonarchX

I installed and tried that program you linked - with, uhm, varying results:

1. it aint working on UV textures (I learned this word today! I'm still proud) - I guess the models are trained on "real" images, so you'd need to train the network on UV maps...

2. any attempt to uspcale not-uv textures (walls etc) resulted in my GPU (1070ti) running out of memory - and I'm not digging into the code to do anything about that here.

so, do you have any textures you want me to try the algorithm on?
Post automatically merged:

PS: So, I let it run on my CPU over night and got upscale images.

However, CPU-bound methods seem to be bottlenecked by my RAM (16GB).

To conclude: unless I find a better or second GPU, or more RAM somehow I can’t do anything about this...

PPS: Why don’t you try it? The page you linked gives an out-of the box method. You only need to download one program (anaconda) and you’re good to go! If you wanna try it, I’m happy to post the install instructions here, it’s very simple (5min install time max)
Last edited:
Damn... Thanks for letting me know! Do they provide any info on the system specs they recommend? GTX 1070 Ti has 8GB of VRAM. Mine has 11GB, but I have same 16GB of RAM. I'll give it a try... Are there command to batch-run it?
Yeah I made myself some batch commands - I can give you all install instructions and bat files.

RAM is only needed if you run the algorithm off of your CPU, otherwise VRAM.
If you think it will fit into my VRAM, then yes, I would appreciate some instructions, 3D Vulkan-based GUI, make sure it can utilize all 12 threads on my CPU (optimize for multi-parallelism), include OpenCL-based live-preview, I want voice commands, instant Chrome-Imgur-upload plugin, and most importantly, once the image generation is complete, make it say "Daaaaamn..." :D.

I think UV textures must be approached one per-frame. So a 8x8 subUV texture should be divided into 64 textures/files and then ESRGANed
@rfuzzo just give the man the tools he needs to make this happen!
To make the tool happen is to make IT happen - same thing. With the type of tool I am thinking of, anybody can make IT happen, anybody who can run it can start generating textures, MonarchX is finally osbolete :D!

Halk Hogan VS ESRGAN - (in Mortal Kombat voice) - "Fight!"
I dunno what you guys are talking about, but here are the install instructions for you @MonarchX to follow and try ESRGAN out for yourself.

1. Download Miniconda ( and run the installer.
2. Update Miniconda: In the start menu, open 'Anaconda Prompt' (it opens a command line window) and type in 'conda update conda'
3. Install Pytorch: In Anaconda Prompt: 'conda install pytorch torchvision -c pytorch'
4. Install Python packages: In Anaconda Propmpt: 'pip install numpy opencv-python'
5. Download ESRGAN (as described in the gitHub Page) - where doesn't matter really

6. Run the upscaling: Open Anaconda Prompt:
- Place a low-res picture in the ./LR folder
- Navigate to the ESRGAN location (with cd - look it up if you don't know how to navigate in command prompt)
- run the test (still in Anaconda Prompt): python models/RRDB_ESRGAN_x4.pth
- Output files are in ./results

as you see, it's all very simple.

PS: "With the type of tool I am thinking of, anybody can make IT happen"
Did you read my first post? exactly not, I am bottlenecked by RAM, VRAM - and (most importantly) the algorithm itself. You need to train it properly! I attach some results for you to judge. here:

And after that, you *need* to try it out yourself, once you done that, we can continue talking - before that, there is nothing to talk about.

Last edited:
Just making sure everything is remaining smooth, congenial, and that everyone remembers to treat each other respectfully. Can't help but feel the heat coming off the last few posts.
Didn’t mean to. I really just meant what I wrote - you should try it for yourself before speculating because it might well be that the tools isn’t the right thing for the occasion after all.

Btw topaz gigapixel has a GUI and allegedly delivers better results maybe that’s something to check out (thanks Olek)