I think it always has had a trial version. Back in the early 2000's my friend and I would mess around with the settings, hit render and then go upstairs to play with lego's or something for a couple of hours, and when it was done it would just be fancy terrain but without details like grass, trees or even rivers I think.
This project generates synthetic computer vision training data. The arxiv paper has more detail including some cool pictures of random creatures it can generate. The images are nice but all of them are nature settings so I assume one would have to supplement this type of data with another data set for training a computer vision model.
Maybe this IS the demoscene of today? I saw more insanely beautiful computer generated pictures in the last couple years than I saw in previous 10, AI or no AI.
This feels like we've got all the pieces of the puzzle to create a reality experience - I'm pretty sure with visuals like this and haptic feedback that your brain will fill in any gaps once you adapt to this given enough time.
You could use this with a VR headset, monitoring heart-beat, temperature and adapt the environment based off the experiencer's desires.
It feels like we're on the precipice of recreating an experience of reality, which may reveal more about our existing reality than we ever expected.
From the homepage it sounds like they've prioritised geometry fidelity for CV research rather than performance:
> Infinigen is optimized for computer vision research, particularly 3D vision. Infinigen does not use bump/normal-maps, full-transparency, or other techniques which fake geometric detail. All fine details of geometry from Infinigen are real, ensuring accurate 3D ground truth.
So I suspect the assets wouldn't be particularly optimised for video games. Perhaps a good starting point though!
Epic did say that you might in some situations forego normalmaps with Nanite and save disk space even though you have super detailed models so it DOES fit in this context.
Also, video games are used to take a high poly model and bake a normalmap corresponding to it on a lower poly model anyway so it might also be used that way. I think Doom 3 was the first game to show the technique?
With nanite normal maps are less required than otherwise because the detail is preserved in the mesh. You could make the argument that micro detail normal maps are still useful but those aren't always generated from the mesh. Especially if they are tiling.
I doubt they prioritized it. To get normal maps you usually first need a high resolution mesh, but then need other steps to get good decimation for lods and normal bake. That's mostly extra work, not alternative work that wasn't prioritized. If by transparency they mean faking aggregates, you also need full geo there before sampling and baking down into planes or some other impostor technique.
This looks rather extremely similar to something that Unreal already natively supports. Here is a demo video from them - https://youtube.com/watch?v=8tBNZhuWMac
It looks great, but I'm missing what's innovative about this? AAA procedural folliage has been done for 20 years, terrain too. Blender has had procedural geo nodes for a long time too. What is so interesting here?
The whole point of this is to generate diverse training data with accurate labels for model training. If you actually want to create nice scenes, use normal blender and some plugins and free online assets.
Blender is pretty approachable. How far are we lowering ourselves to be considered average joe? If it's "a few hours of poking" then sure. But you can create some decent procedural nature stuff with a few free plugins and a weekend or two of learning the basics of the program.
Can we not let creative things take a bit of minimal effort? When we already have tools that make it nearly trivial as is.
true, but that's very easy to do as long as the LLM can create the code, which shouldn't be difficult... idk blender's code lang used (i know godot is very python-like), but most code is nowadays some form of cpp (or it is cpp), so with a few-shot it might be even possible to get a cpp trained llm gen the correct code
Also discussed 30 days ago, 33 comments: https://news.ycombinator.com/item?id=42485423
Watching the video I thought "No Man's Sky as a python lib."
Also terragen but for everything
I miss terragen! What a fun way to waste an afternoon on a rainy day as a kid.
I just looked it up and WOW it has come a long way (and wasn't it free before? Maybe I misremember).
I think it always has had a trial version. Back in the early 2000's my friend and I would mess around with the settings, hit render and then go upstairs to play with lego's or something for a couple of hours, and when it was done it would just be fancy terrain but without details like grass, trees or even rivers I think.
I like the "zero AI" part.
This project generates synthetic computer vision training data. The arxiv paper has more detail including some cool pictures of random creatures it can generate. The images are nice but all of them are nature settings so I assume one would have to supplement this type of data with another data set for training a computer vision model.
The same authors also created Infinigen Indoors[1] to generate indoor scenes for computer vision applications such as robotics & AR.
[1] https://arxiv.org/abs/2406.11824
If only demoscenes were still as prominent today as they used to be. They'd be having a field day with this
Maybe this IS the demoscene of today? I saw more insanely beautiful computer generated pictures in the last couple years than I saw in previous 10, AI or no AI.
This feels like we've got all the pieces of the puzzle to create a reality experience - I'm pretty sure with visuals like this and haptic feedback that your brain will fill in any gaps once you adapt to this given enough time.
You could use this with a VR headset, monitoring heart-beat, temperature and adapt the environment based off the experiencer's desires.
It feels like we're on the precipice of recreating an experience of reality, which may reveal more about our existing reality than we ever expected.
This seems extremely cool. I’m wondering if it can be used to create procedural video game assets.
From the homepage it sounds like they've prioritised geometry fidelity for CV research rather than performance:
> Infinigen is optimized for computer vision research, particularly 3D vision. Infinigen does not use bump/normal-maps, full-transparency, or other techniques which fake geometric detail. All fine details of geometry from Infinigen are real, ensuring accurate 3D ground truth.
So I suspect the assets wouldn't be particularly optimised for video games. Perhaps a good starting point though!
Well, we've come a long way. Look at nanite - it might actually be compatible...
Epic did say that you might in some situations forego normalmaps with Nanite and save disk space even though you have super detailed models so it DOES fit in this context.
Also, video games are used to take a high poly model and bake a normalmap corresponding to it on a lower poly model anyway so it might also be used that way. I think Doom 3 was the first game to show the technique?
With nanite normal maps are less required than otherwise because the detail is preserved in the mesh. You could make the argument that micro detail normal maps are still useful but those aren't always generated from the mesh. Especially if they are tiling.
I doubt they prioritized it. To get normal maps you usually first need a high resolution mesh, but then need other steps to get good decimation for lods and normal bake. That's mostly extra work, not alternative work that wasn't prioritized. If by transparency they mean faking aggregates, you also need full geo there before sampling and baking down into planes or some other impostor technique.
That's actually a fairly ideal fit for nanite meshes.
This looks rather extremely similar to something that Unreal already natively supports. Here is a demo video from them - https://youtube.com/watch?v=8tBNZhuWMac
It looks great, but I'm missing what's innovative about this? AAA procedural folliage has been done for 20 years, terrain too. Blender has had procedural geo nodes for a long time too. What is so interesting here?
Don’t seem to be photorealistic images according to today’s standards
Isn't their logo just the (old) Visual Studio logo?
Source: https://github.com/princeton-vl/infinigen
I was very confused by “math rules only” - as opposed to what? But it seems they don’t think an LLM is maths. Which it absolutely is.
While mathematics are necessary to build LLMs, they are not a kind of math or a distinct branch of mathematics.
Yes I’d agree.
It feels like l-systems on mega steroid, cool
An LLM frontend for this would be amazing.
You could describe the scene you want in normal english and iterate eith conversation.
It could automatically create scenes from novels and poems.
The whole point of this is to generate diverse training data with accurate labels for model training. If you actually want to create nice scenes, use normal blender and some plugins and free online assets.
I was thinking of it as an end-product for use by the average joe rather than a tool for professionals like blender.
Imagine wearing an apple vision and building your memory palace through a verbal conversation with an AI.
Blender is pretty approachable. How far are we lowering ourselves to be considered average joe? If it's "a few hours of poking" then sure. But you can create some decent procedural nature stuff with a few free plugins and a weekend or two of learning the basics of the program.
Can we not let creative things take a bit of minimal effort? When we already have tools that make it nearly trivial as is.
https://www.thegrove3d.com/
https://blendermarket.com/products/quick-landscape
https://blendermarket.com/products/bagaivy
https://forums.unrealengine.com/t/plugin-landscape-combinato...
true, but that's very easy to do as long as the LLM can create the code, which shouldn't be difficult... idk blender's code lang used (i know godot is very python-like), but most code is nowadays some form of cpp (or it is cpp), so with a few-shot it might be even possible to get a cpp trained llm gen the correct code
Is what?
"Infinigen is a procedural generator of 3D scenes".
That's not in the title, where it belongs.
Look at all the sibling posts that are done right:
TabBoo – add random jumpscares to websites you're trying to avoid
Stratoshark, a sibling application to Wireshark
Hunyuan3D 2.0 – High-Resolution 3D Assets Generation
Flame: A small language model for spreadsheet formulas
JReleaser: quick and effortless way to release your project
The infantile downvoting of anyone who calls out useless titles just plays into HN's rep. Knock it off.