The following was preliminary testing. Jump to the most current testing results in this thread.
I am testing various xLODGen settings for "Terrain". I compare the same two areas (snow/no snow). So this is not comprehensive, but gives people some ideas of settings results (on my system) with STEP 0.3.0b. It is a relative compare of different terrain LOD results, all else remaining equal.
skyrimprefs.ini LOD settings:
[LOD] fLODFadeOutMultActors=15.0 fLODFadeOutMultItems=10.0 fLODFadeOutMultObjects=30.0 fLODFadeOutMultSkyCell=1.0000 [TerrainManager] bShowLODInEditor=1 fBlockLevel0Distance=50000 fBlockLevel1Distance=72500 fBlockMaximumDistance=94000 fSplitDistanceMult=4.000 fTreeLoadDistance=72500
Based on my review of sheson's xLODGen thread and documentation, it looks like we should be using best quality and largest textures at LOD4, baking normals for best result (only if normals size ≥ 512 at a given level). Why? Because, in RL, resolution of our environment increases as it gets closer. Also, sheson recommends not tweaking brightness/contrast/gamma without good reason: he suggests lightening/darkening the noise map instead. I have confirmed that changing gamma between 0.97 and 1.10 has no apparent impact in game (data not shown). Lastly, raise steepness should be a good add for LOD8/16/32, since it increases bump effects for mips (probably not necessary at LOD4, due to higher quality textures, and closeness of these textures). I did compare using/not using raise steepness, and I didn't find any really noticeable difference (Lo setting does not include at LOD32, but all others do), so I am including it simply because it sounds rignt "just in case". Compression type does have a significant impact on quality, even for diffuse. BC7 is observably superior to DXT1, particularly with respect to shore-line jaggies around water.
Compare order: Optimal settings (my system, my opinion) > lo settings > hi settings > ultra settings. Links to the settings are also below.
My system is agnostic of these settings. I take no detectable performance hit either way, but I expect that typical PCs will see some difference of more/less significance (system specs in sig).
Optimal Settings - LOD data ~ 5 GB, gen time ~ 45 min
Lo Settings - LOD data ~ 2.5 GB, gen time ~ 25 min
Hi Settings - LOD data ~ 9.5 GB, gen time ~ 45 min
Ultra Settings - LOD data ~ 33.5 GB, gen time ~ 45 min
It would be great if someone with a PC closer to that recommend for STEP and also running STEP would test xLODGen and let us know about performance impact and time to generate.
After getting some feedback from sheson:
- Added Solitude (looking South from near the main gate)
- Increased quality on Lo settings to '10' for LOD16 and LOD32
"optimize unseen of 550 needs a quality setting <= 10 to work its magic in that area for LOD level 32"
- Examined VRAM usage in all scenarios and include a compare of with/wo mipmaps in the 'optimal' settings (apparently, it's not clear if mipmaps are used at all by LOD. sheson confirms that diffuse does not use them, but normals might)
Updated Lo Settings (else they are same as linked previously)
Compare order: Lo > Hi > Ultra > Optimal > Optimal (without mips)
Optimal Setings (no mips)
- FPS stats are largely unaffected by terrain LOD quality on my PC.
- VRAM usage was consistently lower for Lo settings relative to all others and no-mip-Optimal settings (versus Optimnal WITH mips). Not a huge diff, but lower well beyond the +/- 200 MB variation in reported VRAM usage. Optimal (no mips) was comparable to Lo in terms of performance but noticeably better in terms of quality.
- VRAM usage stats are not entirely reliable, because they varied by +/- 200 GB on my PC on any given save-game load once stabilized. This is consistent with how VRAM (and RAM) behave. Algorithms are being employed at any given time in any given system that govern use of RAM. It depends on what the GPU is 'thinking' at the time a save game is loaded: are you loading from a fresh game boot via MO? Did you recently reboot your PC? Are you loading one save game after another? .. it is inconsistent, but not wildly so. My findings are that VRAM usage is a bit less with Lo settings and by NOT enabling mipmaps/raise steepness. Based on the screen compares and VRAM usage, there is no reason to enable mipmaps or "Raise Steepness" on my system.
Simple Summary (for my PC)
Use Optimal Settings (no mips) or some proportionate deviation from that: LOD data ~ 4.2 GB, gen time ~ 36 min
This takes a bit less time to generate LOD, probably decreases VRAM usage to a marginal extent, and results in quality that is all but indistinguishable from Ultra settings. DXT1 compression does not appear to save much besides disk space and costs in quality, so use it on diffuse maps ONLY if disk space is a major issue.