pouët.net

video here : https://www.youtube.com/watch?v=qJcRHrp60t8

navis: while it's true you could probably do a fake version of everything as you describe, it would look worse. i dont want it to look worse. the extra 10-20% of visual quality takes a vastly larger amount of work and technical effort than the fake version. so i have to break with tradition about discussing on pouet, and argue the points. honestly i think you are oversimplifying this due to, well, ignorance. :)

for example, the red pulses, neurons: it's the lighting from emissives & volumetrics from emissives that make it look good - that's what 2023 brings vs 2004. volumetrics give it depth in the way glow does not. you could fake it to some extent with point sources but it's the details of the interactions when it comes from the actual polygons that make it feel natural. youd have to fake it really really well with a lot of detail.

the fractal zoom - if you can do this in the simple way with lodding and make the lodding look perfect, and also make the shadow look perfect and sharp all the way through (the real challenge), and not hide anything under a post effect, i think you actually deserve more technical props than doing it the way i did it, imo - it's probably easier raytracing it. i can give you the model if you want to try it, it's > 1m polys and there are at least 150k clones, you can see many of them when zoomed out. if it was lower poly, when you zoom in you would see the low polys of the higher levels when they get huge. it has to be totally seamless. maybe the shortfall in that scene is that it doesnt make it obvious enough visually about the counts. but it's a joke anyway, a reference to dope/complex. 7800 phong shaded polys.. :)

lights hanging from the floor: it's the detail (the way the light affects the character, the diffuse light interaction) that matters, really. if you fake it in the way you describe it looks like we're back to 10+ years ago.

voronoi cuts on text: if you precomputed it would be a lot of models, because the text animates constantly. also the voronoi cut pattern remains constant. that would be quite painful to achieve in 3d software as well. so you could fake it, but it would be really annoying to do it.

foliage: the challenge is about the poly count and the lighting quality. the shadows are ray traced and that gives it a much higher quality than shadow maps would, but at the same time makes the motion difficult because of the constant data structure rebuilds.

smoke: you could bake it, but the caches are massive. 1gig would be a distant memory. or i guess you could just render it out and comp it, but that really is a fake too far isnt it?

it was important for me personally to try and get it to look as good as possible. to have a render that feels good, even if the reasons why are not totally in your face, but just give an impression of "looks good". be it a correct reflection, an emissive interacting with the volumetrics, shadows in the foliage giving them the right sense of depth & volume as they move, the sharpness & cleanness of the render, etc. small things that add an extra few %.

also to be honest the process matters more to me. i was not really interested in spending my time working out how to fake everything, i wanted to achieve the result in a way that made me enjoy it. it's been 5 years since i last did a demo (and the first one i did on my own for even longer - no design superstars on this one which is probably why the direction is worse), and my interests have definitely changed - it's more about the process itself nowadays. doing it "for real" is a more enjoyable process because it's creative and iterative and you can change things easily, whereas the fakes lead you down a garden path with no other ways to turn at the end.
i guess it could be a fun challenge one day to do something that really is heavily faked and do it well, but just don't make it look cheap (as the fakes you describe would be).

but i must admit that it runs too slow with less VRAM (outside of compo machine) and it takes too long to load, and some stuff is a bit janky (ran out of time).
the main reason for the slowdown is VRAM swapping - the individual scenes actually run fine on a 3070 (i was on my laptop for last few days with that GPU, in fact). but when you put it together the VRAM use is a disaster, and it's that swapping that causes the slowdown. but i ran out of time and the compo machine had 48 gig, so the sacrifice was made. may be able to fix that.

the annoying thing for me is the load time, this is only a problem when loading from the compiled version, in the tool it's much faster. something is broken there and i didnt get time to diagnose it (the player only worked at all the day of the deadline..). it should actually load much more quickly than our older demos. this bothers me, may try to fix it.

the filesize..? it's 2023, guys. but also, the reason it's so large is that it was engineered to load quickly and run well, at the expense of size. for example there are like 200mb+ of compiled shaders here, compiling those at load time would be a long wait (why 200mb you say? well ^^ go back to the "visual detail" above.). the 3d is converted to a format that's fast to load not small (and again, focusing on quality and not saving bytes).
so in that respect i was not successful, it's a valid criticism, because it doesn't load quickly and it runs too slow without a lot of VRAM - that is annoying. but it ran perfectly on the compo machine and was totally out of time, so had to say fuck it for the compo itself, i had already pushed my deadline luck to the limit.

T-shirt design sorted.