Avalon local LLM

[Pocster]

Been bugging me for a while. Whilst the terrain is good it's not good enough (for me). REALLY want to ray march , notoriously difficult and gpu expensive. Doing ok ish. But we have no cliffs, valleys etc - because erosion makes them. So we need to simulate erosion accelerated and 'carve' through the landscape. Decided to compromise with realtime procedural infinite world and go finite (but loops), still procedurally generated of course. Trying Opus 5 as yesterday with 4.8 all attempts at erosion and looking realistic failed. Changed from webgl to webgpu aswell - means everything previous is no compatible. So a big re-write is due at some point.

But this is difficult. Everytime we attempt erosion we get some weird mess on a world (doesn't take many spins at random to find one). So I've set Claude up to render them itself, do a top down view and not talk to me until its gone through 50 seeds (pre determined for testing) and 50 random and they all look realistic. Either it will do it or I'll run out of tokens!

 

Edited yesterday at 13:29 by Pocster

[Pocster]

Been a slog. Up'd opus 5 to max (no idea if that truly benefitted tbh). Perseverance pays off with these things though. Near photo realistic landscapes now - eclipses previous effort by some margin.

But nothing to show yet until we beat my personal PITA ..... WATER.

Such a difficult thing to get right; always a compromise on visuals against performance. Usually looks cheap or just too "gamey". So like the terrain we are going for realism. To achieve that I decided we need to model how/why/where volumes of water form. My logic being if its physically correct then we are half way to it looking correct. Usually these things are cheats like most 'effects'. So we use what we have plenty of... ram. For shadows for example to save a few computes per pixel the entire landscape shadow map is pre computed (around 8gb) - it can be optimised later size wise. Doing just this knocked 5% of render time. I can think of similar such tricks for water.

[Pocster]

water now is staggeringly good! Proper beer/lamb for depth, waves, lapping, water's a bugger because the same water looks different depending on your view. Look at your feet see the seabed, look across it and it can be almost dark. Managed to get that working successfully.

 

[Gone West]

57 minutes ago, Pocster said:

water now is staggeringly good!

The second picture is good except for the regularity of the wavelets. There is something odd about the first picture, but I can't really put my finger on it. Maybe the dark area in the distance shouldn't go right up to the beach, perhaps there should be a thin strip of blue by the sand.

[Gone West]

Oh dear, perhaps I shouldn't have gone there.

[Pocster]

1 minute ago, Gone West said:

The second picture is good except for the regularity of the wavelets. There is something odd about the first picture, but I can't really put my finger on it. Maybe the dark area in the distance shouldn't go right up to the beach, perhaps there should be a thin strip of blue by the sand.

Yeah the waves do have a mathematical view. Easy to see when static, less so when moving. Procedural generation will do that!. 

[Pocster]

3 minutes ago, Gone West said:

Oh dear, perhaps I shouldn't have gone there.

I'll do a video at some point. Water from a certain angle with certain lighting looks 100% convincing. Static's can't show that. Regarding your 2nd comment - not sure exactly what you mean and again depends on camera angle. What's the dogs here is zero lod, zero geo morphing, zero z shadow fighting. So terrain no matter how far you can see looks the bollocks. Focus on 1 peak in the distance - move towards it; perfect!. Also remember Work In Progress!__dbg.marcher.u.waveEarlyExit.value = 0

[Gone West]

1 minute ago, Pocster said:

What's the dogs here is zero lod, zero geo morphing, zero z shadow fighting.

If only I knew what you were talking about. The last picture looks great by the way.

[Pocster]

2 minutes ago, Gone West said:

If only I knew what you were talking about. The last picture looks great by the way.

All the things as an ex games programmer I notice and hate. But with finite compute and finite ram compromises must be made. Give me and my new best friend Claude (I've not even sworn at it yet! , honestly not once! , whilst CHAT , well (expletive deleted) me ... I couldn't stop swearing ). The point you made is valid "looks wrong, but not sure why" - I do that a lot. I have to explain what IS wrong - usually in a technical term. So you find things visually 'wrong' thats fine. I just grabbed some random screenshots tbh .

[Gone West]

5 minutes ago, Pocster said:

So you find things visually 'wrong' thats fine.

 

Looking at it more, I think it looks odd because the blue water travels around the edge of the beach, but only for a short distance, before turning dark by the beach. Why in reality wouldn't the blue continue all the way along the beach? There is nothing to indicate that there should be shadow in that area.

 

[Pocster]

19 minutes ago, Gone West said:

 

Looking at it more, I think it looks odd because the blue water travels around the edge of the beach, but only for a short distance, before turning dark by the beach. Why in reality wouldn't the blue continue all the way along the beach? There is nothing to indicate that there should be shadow in that area.

 

Te-he! As I said water looks different from different angles. So near us (the camera) you see the turquoise and seabed further you would not at this angle; i.e. you cant see the nice blue because our angle is so low - we are looking over it. If I pulled the camera up into the air and looked down at that spot it's there. So looking at water and looking across water are very different views - it's something through the previous 5 versions (oh my!) I just couldn't get correct. look at piccy 1; water we are near is 'clear' (depends on a LOT of maths, sun angle, sea bed, refraction, specular etc etc etc etc ), but look across at the mountains, a reflection and is dark. Beer-lambert (remember you started this science/maths argument! 🙂 ). But I'll let my mate Claude answer....

 

"

The blue stops because of viewing angle, not shadow. What you're seeing as turquoise is light that passed through the surface, reflected off sunlit sand, and came back out — and how much survives depends on how far it travelled through water, not how deep the water is. Directly in front of the camera you're looking down into it, so a metre of depth costs about two metres of travel and the sand shows through. Along the beach receding to the right you're looking almost flat across the surface, and the same metre of depth costs a hundred times more travel. Nothing survives that, so it goes dark.

That much is real: at grazing angles you genuinely do stop seeing into water, and a photo taken from that low across a lake shows a mirror at the far shore rather than turquoise.

Where ours overdoes it is that the surface doesn't yet bend the view ray. Refraction puts a hard ceiling on the underwater path — about 1.5× the depth, from any viewpoint — so real shallows stay bright the whole length of a beach. Without that ceiling the path grows without limit as the view flattens, and the band dies far earlier than it should. Separately, what replaces it should be a bright mirror of the horizon sky; ours fades reflections to black near horizontal, which is why that stretch reads dark instead of silver."

[Gone West]

2 minutes ago, Pocster said:

But I'll let my mate Claude answer....

I wouldn't argue with Claude!!

[Pocster]

6 minutes ago, Gone West said:

I wouldn't argue with Claude!!

I do. Sometimes it's wrong. But trust me accuracy is paramount to realism - I'm on it!

You've got bedtime reading of Beer-Lambert and quaternion maths. You wont like either of those trust me!

Edited 6 hours ago by Pocster