Jump to content


  • Posts

  • Joined

  • Last visited

  • Days Won


vozka last won the day on December 1 2021

vozka had the most liked content!


74 Excellent

1 Follower

Recent Profile Visitors

The recent visitors block is disabled and is not being shown to other users.

  1. Added some chipped paint to the painted version. I'm the type of person who hates doing things by hand so I spent about 5x more time learning how to do it procedurally (surprisingly there are a few catches), but I'm pretty happy with it and I think that makes this one worn enough.
  2. If one out of ten people notice, that's good enough for me, the only issue is if I'll be able to make it low-poly enough to not strain the engine. I tried making it low-poly and bake normals in a quick & dirty way, so there are errors and I'll have to try a couple times, but so far I got the whole handle down to 170 tris. Seems like with a shadow mesh, which I guess could be half that, it would be good enough.
  3. Grunge and distressing is definitely something I have some trouble with, but you're right. Regarding the inlay, good point - I didn't really do that because this is still PBR with metalness and everything and the materials will have to be done slightly differently for TDM anyway. While this is flattering, it's not that difficult to get a semi-photorealistic render in a path-tracing renderer using physically based materials and models with a virtually unlimited number of polygons. Adapting that into the game is another matter I created two variants. First is a black paint (again, no distressing yet) that imo looks a bit better, but if I recall correctly it's possible to have models with more than one set of materials, so it would be best to have both. Second is a new handle what apparently was semi-common in the victorian times and I thought it was pretty funny. No idea how it will look after low-poly retopo and baking though.
  4. Awesome, that would be great. Regarding the separate models, that's what I assumed, so that's how I did it. I'll work on it a bit more first.
  5. Well if this is deemed good enough, I'm probably going to need somebody to help with the interactivity anyway - door opening and the lock/moving the handle. I'm not a map maker and I'm not sure I'll have the time to learn that (it's harder to justify than practicing modelling, which I need for other things too). Providing a version without the keyhole would not be a problem for me of course. I have an alternative material and handle in mind as well, it would be nice to have more than one option.
  6. Attempting to make a model of a safe for fun and practice. So far it's just hi-poly with PBR materials. My modelling skills are quite narrow and this is outside their scope, so I'm looking for criticism, mostly: - Is the overall style acceptable, regarding apparent time period etc? (apart from the text in the logo, that's a placeholder and I imagine I'd use some decorative victorian font plus some actual believable names) - Would the quality be good enough like this? - Do you have suggestions about how to improve it?
  7. Yes. Not as good as Dall-E at this moment (but will improve soon, several improved versions are already in works or done and will be gradually released), but free and open-source.
  8. An article about using Stable Diffusion for image compression - storing the data necessary for the neural network to recreate the image in a way that would be more efficient than using standard image compression algorithms like jpeg or webp: https://matthias-buehlmann.medium.com/stable-diffusion-based-image-compresssion-6f1f0a399202 It's just a proof of concept since the quality of the model is not there yet (and the speed would be impractical), it can't do images with several faces very well for example. But it's very interesting because it has completely different tradeoffs than normal lossy compression algorithms - it's not blocky, it doesn't produce blur or color bleed, in fact quite often it actually keeps the overall character of the original image almost perfectly, down to grain levels of a photograph. But it changes the content of the image, which standard compression does not do.
  9. Also most missions have so few ratings that the average is kind of random. If you're just looking for some order in which to play the missions than this is not a bad one, but it's not really worst the best and don't take the ratings too seriously.
  10. This is already possible to a degree. It takes two things: first you use the language model of the AI to "interrogate" the image and find what text description would lead to generate something close to it. Then you basically use the diffusion algorithm which normally generates images from noise in reverse, to gradually generate noise from an image. This gives you the seed and description, and the result is often very close to the original image. I was playing with it yesterday because it then allows you to change some details without changing the whole image, and that allows you to take the portraits of your women friends and transform them into being old and unattractive, which I found all of them universally appreciate and find kind and funny. The early implementation in Stable Diffusion is just a test and it's quite imperfect (it produces noisy images with overblown colours among other things), plus it's limited by the quality of normal SD image generation, but as a proof of concept it obviously works and will probably get better soon. However I don't see this happening any time soon. The AIs work with bitmaps, are trained on a specific image size (with Stable Diffusion it's 512 x 512 px at the moment) and making anything significantly smaller or larger completely breaks the process. Various AI upsampling algorithms exist, but they never work as well as straight up generating the image in the resolution that the neural net was optimized for. And I don't know about any practical solutions to this yet.
  11. This is offtopic, but you just reminded me of a great memory. I was always into underground music culture, mostly metal as a teenager, but later also electronic music. A few months after I moved to a different city for university I discovered a tiny underground (literally) private bar, basically a rave speakeasy. No license, entrance with a camera, located in what used to be a wine cellar and a later a PC gaming arcade (which left very strong air conditioning there, so the weed only smelled outside, not inside). Each weekday it was open from about 6-7 pm to about 4 am, and almost every time there were one or more DJs playing more or less underground electronic music, for free, just for fun. Sometimes doing live broadcasts over the internet, before that became a common thing. I still remember the first time I came there, walked down a narrow set of stairs and entered the dark main room with a couple tables with benches and about 10 different CRT monitors placed in random places (including under the benches), all playing Electric Sheep animations. At the time it felt like something from The Matrix. The place is still kind of operating many years later, but it got handed over to guys who care more about getting drunk everyday than about music or technology, so the magic is long gone.
  12. Neural networks can and do forget things during training, and their memory is very imprecise. And we're not entirely sure why that is, as far as I know. But it's one of the basic principles without which the whole concept wouldn't work. It is true that a neural network cannot have a sudden flash of insight, or more precisely it cannot realize that it just had a sudden flash of insight. Artificial general intelligence does not yet exist and is probably a couple decades away, and specialized AI is not self-aware. It can certainly act as a flash of insight for a human that uses it to create something unexpected though.
  13. You make some good points, but I personally think it's similarly naive to think that we cannot eventually get to a point, where an artificial intelligence produces results so similar to a human in some areas, that the fact that it functions differently becomes irrelevant. In fact I'd say that current neural networks are a big step forward both in terms of the results they produce and in terms of how similarly they work to a brain. We know that artificial "neurons" are a distant abstraction of how a brain works, but a neural net is not just a normal mathematical algorithm. Yes, it's "just" matrix multiplications, but at the same time the complexity of the computation is so far beyond what any human could analyze or design that it's a blackbox. That in particular is similar to a real brain. The output of a neural network is deterministic just like with a "normal" algorithm - two identical inputs give you two identical outputs. With the case of image generation algorithms part of the input is noise (in the shape of the final image), which dictates how will the final image look. Change the noise very slightly and the output can be completely different. Since we want the results to be reproducible, we use a pseudorandom noise generator initialized by a seed number. Same seed will produce the same noise pattern and therefore the same resulting image. What if we instead use a noise generator that is truly stochastic? We are able to do that (using specialized hardware). The algorithm itself would still be deterministic, but it would be unpredictable because a part of the input would be unpredictable. Would that be "organic" enough? And what if we stop computing the neural network on a GPU and instead use a chip that does analog matrix multiplication (already exists, although not powerful enough for this particular usecase, yet), which brings a stochastic element into the computation itself? The computation itself is no longer deterministic. You know what will roughly happen, but cannot predict the exact result even if you know the input. Would that be organic enough? It is certainly a very different thing now from how any normal computations work.
  14. This will undoubtedly happen with large parts of digital art. But there's a ton of art in the physical world that is not in danger. Even oil paintings - you can print digital art on canvas, but it won't be able to mimic some specific painting techniques that create a physical structure on the canvas, or specific types of pigment. Or various types of combined media. Then there are other even more physical media like sculptures, that's an infinite world. Honestly from the various art school graduates/teachers that I know none of them are threatened by this because none of them ever did normal digital art like anything an AI can produce.
  15. That's the beauty of it being open-source. An unofficial AMD port already exists (although I have no idea how well it works) and people did several optimizations that reduced the RAM necessary to run it down to about a half (with almost the same computation speed, less than half with reduced speed). My results are just a proof of concept and not very useful for game assets, but I was pretty happy with them. I made a very simple 3D render with unnaturally clean PBR materials, and then added some dust and dirt on them using Stable Diffusion. I had to overlay some black noise texture over the clean materials because the AI needs something dirt-like to work with, it will not generate dirt on something that is completely clean unless you let it deviate from the base imagea lot, and in that case it tends to change the shape of items and distort things. I also had to render the bench and garbage bin separately (but keep their shadows in), the AI tended to change them too much. The images are 832 x 832 px in size because that's the maximum that currently fits into my 6 GB of VRAM (I'm not using the latest optimizations), but when doing something like generating noise (which doesn't need much continuity), you can easily just split a large render into smaller tiles and do it tile by tile. Before AI dirt: After AI dirt:
  • Create New...