SMIM 4K - Cleaned and Upscaled Textures for Static Mesh Improvement Mod (by Brumbek - XilaMonstrr)

[Mousetick]

Discussion topic:
SMIM 4K - Cleaned and Upscaled Textures for Static Mesh Improvement Mod by XilaMonstrr
Wiki Link

---

Complete Texture Replacer for the great SMIM, with the same FOMOD options as the original. Textures have been cleaned and upscaled to 4K with modern compression. Plus performance options for Potato PCs.

Spoiler

Quote

The original SMIM textures were a mix of BC1 and BC3 compression. The resolutions were a mix of 1K, 2K and 4K (if using the High-Res options in the custom install, otherwise they're lower--including in the AIO installations) plus some 256 and 512. I have cleaned all files of compression artifacts caused by these older compression formats, and then upscaled them all by 4x, to a maximum resolution of 4K. Then I converted them to BC7 compression and generated full mipmaps. I cleaned and upscaled them using Cupscale/ESRGAN and then finalized with Octagon. 

Notes:

• Judging by the comments on similar mods, I understand Tech is not a big fan of this kind of mass-processed AI upgrade, and for good reasons. I find they're often too sharpened and darkened. I'm still putting this mod up for suggestion and recommendation, as I think its quality is very good, it improves upon over the original, and there is no equivalent out there (that I know of).
• Similar mods by the same MA relevant to STEP guide:
  • Deadly Spell Impacts 4K - Cleaned and Upscaled Textures
  • Ruins Clutter Improved 4K - Cleaned and Upscaled Textures
  • CoverKhajiits - Face Textures 4K SE (already included in STEP guide)
  • Detailed Rugs 4K-2K - Cleaned and Upscaled Textures
  • Book Covers Skyrim REDUX 4K-2K
  • Book Covers Skyrim - Lost Library REDUX 4K-2K
• I've been using all the above (in 2K or 2K/1K for Potatoes) ever since they were released.

[z929669]

It looks like the upscale method used is 'good', but I only question the conversion method used. If each texture was upscaled and QA'd by a human, then it's fine, but it hundreds of textures were batch processed and batch upscaled (through CAO, fe), then there could be lots of underlying issues without proper QA.

I think this and the other MA's mods listed are all worthy of consideration if some level of a hands-on approach was taken to verify that the treatments applied were not one-size-fits-all (unless it was confirmed that this is actually true). If any texture looks off or has bad mips or alpha-to-coverage, etc., that's a sign that they were batch processed with limited QA (potentially fixing some things and introducing new issues elsewhere).

These are the kinds of mods that need testing or time for any bugs to crop up on Nexus if testing is too daunting.

---

PS: at a glance of the Nexus page, this one does look good. Released months ago, no bugs, lots of downloads and endorsements, etc. So if it is true, then all of the other mods become more interesting as well.

[TechAngel85]

This author is okay and they do okay work, but they do batch work for their stuff as well. It seems they're on a mission to just upscale all the mods on Nexus. lol

Anywho, this one is no different from the others. It's going to have some good and some not. I'll pick on it for a moment...
The first shot on Nexus looks like plastic, not wood. It, all the wood texture shots, and the coal pile has lost any purposeful noise in their textures, the upscale has given the coal an orange peel texture, that Normal example has new vertical lines in it caused by too harsh of processing when there is heavier artifacting in the original texture... 

That's just me picking on the example we've been given. Yeah, they look nice, but there are still going to be issues due to the nature of batch processing. Anyone can do find a few good models and just batch a bunch of texture through them. It's super simple to do as the programs do all the labor. What sets upscales apart is what is done to reach the models used and what is completed after the upscaling is completed. I'm not going to go into what I consider to be a true upscale, but my process of upscaling can easily take 1-2hrs to complete per texture.; mainly due to everything extra I do, manually, post-upscale.

[z929669]

After looking more closely, I can see some of the 'plasticy' look in a couple of the examples as described, but overall, they appear much more "in focus" than the originals ... so in game, it will have the effect of looking more crisp and higher res as intended, IMO. As long as nothing is really broken, I think it's fine and will seem like a net plus in game (which is likely why there are no bugs reported, and the mod is getting used and endorsed).

---

TLDR;

I don's see any issue with the normal though. Normal 'noise' is more forgiving and sometimes has a nice effect ... to me, the normal just looks like it has more wood striations than the original and will likely be a marginal improvement in game even with the original wood texture.

But I would never upscale normals anyway. I would create new ones based on the upscaled diffuse.

On a side note: Upscaling anything is extrapolative and involves adding missing information using a Bayesian sort of approach. It can never be as good as the original when scaled to the original resolution (one of the mip levels of the upscaled texture is a direct compare to judge how well the upscale performed ... you should find that information is actually lost).

This is the one that emphasized the general limitation of upscaling, IMO. The right looks very much like plastic wood (to me), while the left just looks like a picture of wood slightly out of focus:

This is because the 'noise' of the original is rather consistent, and it winds up looking cartoon-ish when upscaled without adding new noise to break up the lack of variation in the minutia.

All that said, I agree that using the deliberative, non-batch approach is always going to result in a better outcome if for no other reason than the custom tailoring to the use case of the end result befitting the original.

[TechAngel85]

12 minutes ago, z929669 said:

This is the one that emphasized the general limitation of upscaling, IMO. The right looks very much like plastic wood (to me), while the left just looks like a picture of wood slightly out of focus:

This is because the 'noise' of the original is rather consistent, and it winds up looking cartoon-ish when upscaled without adding new noise to break up the lack of variation in the minutia.

You can get that looking a lot better by using a less harsh decompression model and then using a different model for the upscale. This is a good example of him using a model that mainly just sharpens on everything and why batch processing is never going to completely work.

I'm still working on my version of upscales for vanilla here and there so all things upscaling is still fresh in my mind.