[2.13] AI vision uses different light estimate system

[stgatilov]

Detection of player, bodies, ropes, etc. by AI depends a lot on how well-lit the object is.

The player case is the most important one. It uses its own system to compute lightgem value, which renders additional views on GPU. Even with many improvements over the years, this approach costs a lot in performance, so it cannot be used for all objects which AIs can see.

That's why there is another system for computing light values for the other objects.
Since the engine was closed source when TDM was originally created, this system is very approximate. It takes into account light intensity and size, but mostly ignores the other factors like shadows, light projection/falloff textures, light texture transforms, blinking lights, lights with some material stages disabled etc.

A new light estimate system has been implemented, and it is going to be default in 2.13. In fact, it has already been default since dev17104-10844.
As far as I know, it accurately takes into account all the properties that our lights have.
It has its own issues of course. It relies on a few samples on the object surface, and it is rather performance-intensive because it casts shadow rays all other the world. I had to distribute its work over several frames to reduce performance cost, so now it takes some delay to compute the proper value (which can also cause problems sometimes).

Here is how it works on a test map:

What it means for us?
On the good side, you should no longer have a problem when you put a corpse into a dark spot under a table but guards still detect it easily as if the table did not exist.
On the bad side, the new light values are inherently different, so guards might detect objects in situations where they did not detect them before or vice versa. Also the delay can probably cause some issues e.g. guards acting upon an obsolete value which is updated just afterwards.

The effort is tracked in 6546.

There are several cvars about this:

• g_lightQuotientAlgo: value = 0 uses the old system, value = 1 uses the new system.
• g_showLIghtQuotient: enables debug visualization like in the video above.
• g_les*: lots of internal parameters to tweak the system (you should not mess with them).

 

[datiswous]

Interesting. Maybe this fixes the issue in fm Who watches the watcher where you knock out someone, put the body in a pitch black shadow but guards see the body from far away.

Edited January 5, 2025 by datiswous

[I.C.H.I.]

11 hours ago, datiswous said:

Maybe this fixes the issue in fm Who watches the watcher where you knock out someone, put the body in a pitch black shadow but guards see the body from far away.

I thought it was by design?

BTW the only FM that gave me real troubles with hiding bodies was Vengeance for a Thief 3: The Art of Revenge, it just felt plain broken because of its poorly set lighting.

[datiswous]

On 1/5/2025 at 4:20 PM, stgatilov said:

On the good side, you should no longer have a problem when you put a corpse into a dark spot under a table but guards still detect it easily as if the table did not exist.

So why is it that in 2.12 and earlier putting a body in a dark corner often does actually work to hide it?

Edited March 27, 2025 by datiswous

[chakkman]

I had issues with guards finding bodies in pitch black spots in the past...