Blog - Home

Comparing graphics presets and fidelity levels in Half-Life: Alyx

16/09/26 05:16

This was originally going to be a post about the graphics settings in Half-Life: Alyx - what they do, and the visual and performance impact they have. The more I researched, the more complicated things got. I won't go into too much detail in this post, but the basics are as follows.

HL:A has dynamic render resolution, referred to as the fidelity level (nothing to do with the fidelity setting in the in-game menu). It will turn the resolution up or down to try to maintain the desired framerate (the refresh rate of your VR headset, with some accounting for overhead). This causes some weird effects on how the game looks at different settings. If you turn the settings up, the game looks better, but gets harder to run - so the fidelity level lowers to maintain framerate, and the game actually looks worse. Vice-versa for turning the setttings down.

On top of this, some of the settings are not well understood. The in-game menu tries to explain the performance impact for some of them, but it's not super clear what they actually do. In the first draft, I started trying to explain the settings, but I couldn't actually find good sources for all of them. So I decided to find out for myself in-game. This turned out to be pretty tricky, and so far I've only been partially successful.

First of all, I needed a way to take comparison screenshots. As it's a VR game, it's hard to position the camera statically, so sequential shots can be compared by flicking through them. I considered placing the headset on a table, but this is an unreliable workaround, so I dug into the console commands (with the help of this gist) and found a solution:

1. Add -console -vconsole to the games launch parameters.
2. Launch the game, adjust graphics settings to the desired preset. Load into the first level, take off the VR headset.
3. Open the console and run sv_cheats 1 - needed for a couple of the next commands.
4. vr_perf_hud 1 to enable the performance HUD.
5. vr_fidelity_level_auto 0 to disable the game automatically changing the fidelity level. This concludes the setup portion.
6. vrfly_enable 1 to decouple the game camera from VR tracking.
7. setpos -270 1930 210;setang 5 75 0 to position the camera on the balcony in the opening level (vrfly ensures these coordinates are relative to the map rather than VR tracking)
8. host_timescale 0 to freeze the game world (so NPCs and such don't move between shots).
9. Set a fidelity level, e.g, vr_fidelity_level 3
10. Take a screenshot of the spectator view with jpeg_screenshot

I repeated steps 9-10 for each of the 9 fidelity levls (0-8) at each graphics preset, for a total of 39 screenshots. However, there's a few problems with this method:

• host_timescale 0 doesn't fully freeze the game world. NPCs still move, just very slowly (one frame of animation every couple of seconds?), and some things like particle effects are unaffected and keep moving.
• I couldn't find a way to change graphics settings through console commands. Some settings seems to use multiple cvars, others don't even use cvars at all. So the only way to move on to the next preset was to reset host_timescale, disable vrfly_enable, put the headset back on, change the settings, go back to the main menu and load back into the level (to ensure the settings were applied), take the headset off, re-enable vrfly and set host_timescale back to 0, re-position the camera back to the same co-ordinates, and then start taking screenshots. Since this all took a variable amount of time, different NPCs are visible or have moved around between comparisons of different graphics presets.
• jpeg_screenshot is pretty low quality - 1080p, and compressed. There is a png_screenshot command, but it didn't work for me - just produced a black image. Taking a screenshot in SteamVR by pressing the trigger and menu button is even worse, since it screenshots the steamVR preview window rather than the HL:A spectator window. I can't really take a screenshot with F12/snipping tool, as the HL:A window steals focus, and with vrfly, the mouse causes the camera to move.
• I'm not 100% sure that returning to the menu and re-loading the level actually fully applies the graphics settings - it may require re-starting the game each time. However, I didn't realise the issue until after I had already taken the 39 screenshots.
• Expanding on the last point, doing this for anything more than the presets would get even more tedious fast. There are 6 graphics settings with low-to-ultra options, and two on/off options. (6*4)*4 = 48 graphics configurations - each requiring the song and dance described above, and possibly a game restart. Multiply that by the eight fidelity levels and I'd need to take 384 screenshots. Double it again for each new area/level. Due to the JPG compression of the screenshots, the differences would barely be visible between most of them. The differences in the position of NPCs/effects from applying host_timescale 0 at different times after the level loads would drown out the graphical differences in most case.

Despite the issues, I've thrown together a page where you can compare the screenshots. Up/down cycles through the presets, while left/right cycles the fidelity level. On desktop, you can use the arrow keys. It requires JavaScript, and full disclosure, was vibe-coded by ChatGPT because I'm feeling pretty defeated at this point and did not want to wrestle with JavaScript (I am not a programmer) at 4am. I hate to subject anyone to AI slop, but if it weren't for the clanker making it for me in ten seconds flat, you would not be reading this post. That was the only use of AI in this whole project. I am trying to quit. Sorry.

As you'll likely notice, there's really not much difference between the settings. I suspect this is because:

• Low image quality.
• The presets don't change the texture quality (it's always set to high).
• Re-loading the level might not have been enough to get the settings to fully apply.
• The opening balcony is ridiculously optimized and not actually a good test for graphics.

The only difference I can spot myself is the shadows of the objects on the table, on the right side of the image. What a lot of work for so little result.

If I can find a way to drastically improve the quality of the screenshots, and reduce the friction/tediousness of changing graphics settings, I may re-visit this project. If you know a better way, please get in touch - see the homepage for contact details.