https://microsoft.github.io/TRELLIS.2/
>>1023641the market for disfigured roman statues might be cornered indeed...
>>1023641not bad for just one picture, the fingers are fucked and the fist too, people have been using photogrammetry years and this is basically the same shit.
>>1023646>fingers are fucked and the fist The wing isn't solved right and neither thie right arm
>>1023641>>1023646Now show us the mesh/topology and the UVs and image textures.
>>1023665>show us the mesh/topology and the UVs and image textures.NTA but why? not everything has to be game assets, anyhow you have mesh generators that make clean topology too
>>1023641>100k minimum polys for this and obviously the geometry is completely melted and unusablePartpacker is the only one which impressed me. That's the only solution - making the AI separate the parts so it's actually usable now nothing is unless for an offline render in the distance even then it's shaky and no one in a serious production will put melted 100k poly assets in the distance.
>>1023641>it REALLY is over...what the fuck does that mean?it's just a new tool
>>1023641I tried; I really did but the results are awful.
>>1023806
>>1023641its over... in 2-5 years maybe
>>1023641I tried it and it’s the same/worse then that Hunyan Chinese model from 9 months ago>All details melty>Badly applied texture, makes it impossible to even decimate>Terrible baked shadows that look bad even in unlit game>0 progress on topology>While mesh melted together, something previous models tried to solve already>Faces still look fucked upIf you showed me this model year ago it would look like small step forward in the state of the art. But now it looks like they went backwards/are stagnating extremely hard. To me it looks like proof that we will need some brand new innovation to solve the current problem, kinda like how image gen got better and more controllable once they started using transformers for images, making multimodal LLMs rather then using diffusion models.
>>1023862AI sucks at these really precise patterns & strict angles. it doesn't get it. the result will always be tessellated. the only solution is to not bother with this point cloud type stuff but just make it build it from scratch via proper box modelling or as close as AI can get to proper box modelling. I'm guessing there's not enough data for it through and there's no real way to scrape it even if you dump it from a bunch of games the result is usually not quite what the artist had in mind because engines convert to tris.
>>1023844> in 5 years the 60 year old algorithm will changelmao, no.
>>1023875I also think there probably never will be, mainly because nobody actually wants it. If you spend some time in the AI discussion sphere where they talk about the latest models and what the lead scientists are saying, you realize that none of the big labs is actually going after art, they are after bigger goals and art is just in their way. Google deepmind for example does the whole Genie 3 and AI agents playing videogames not because they want to automate gaming, but because they want to help robots get thought on data that does not exist, artificial environments. For that all you really need are these point cloud colorful blobs we have today, making good low poly models would not help in the slightest. Furthermore if you also ask one of the AI fags if they want to automate something like 3D modeling they will just tell you that 3D modeling is redundant anyway thanks to video models. Why do you need 3D models when you can just ask AI to animate any image and current point cloud models are sufficient enough to make you 3D printer files.Basically AIfags are too distracted to tackle the problem of 3D modeling, especially since the only use of 3D modeling with good topology would be literally useless, oh we automated those few 3D artist jobs in game development, thats all it would achieve because other tasks can be done with video models anyway. Oh and before you mention image gen as useless, you are kind of right, but again, all of the big labs now develop image gens mostly because they believe in "multimodel AI" basically one AI system that is capable of everything and can generalize in everything, which is what google does.
>>1023641>get a 2d image>use a 9999 ram consuming "ai" model to make it 3d>so you can render it in 2d againwhats the point?>>1024001i feel like all the jebs in visual arts industry that could be replaced by AI are already have been:stockimage websites, ads, porn, propaganda. that's it. this is the extent of application of AI in our current industry. Its kinda like how mcdonalds could just make their shit in factories and ship it frozen to restaurants but they decided to not do that coz having a dude there heat up and lego assemble some frozen food is better for sake of public image. And this is such a lame ass poetic "worst of all worlds" ending to the whole AI in arts story that it just has to be true. You don get posthuman AI creativity revolution you dont get socialist optinal jobs UBI stuff you dont get the butlean jihad what you get is broken dreams and bullshit jobs.
>>1024016Yeah, the AI industry is forced into tackling bigger problems, not because they have the capacity to solve healthcare or engineering, but because the valuation of the AI industry is bigger then one third of US GDP or 1/10th of the global GDP. OpenAI alone has bigger valuation then the entire gaming industry combined, and more then 5 times the entire film industry. Art for them at this point is too small fish, this is why they are promising UBI utopia where AI does literally everything from curing cancer and discovering new laws of physics to wiping your asshole every time you leave a toilet, because anything less would make all these companies look overvalued. The bubble is worse then any other in human history, the only thing protecting it from popping is the fact that all big AI companies have bigger company selling actual products to foot the AI bill (Google own most of the internet, Microsoft most of operating systems, NVIDIA makes the best hardware, Amazon most of retail) and because they keep promising that the next AI breakthrough will be white collar bloodbath and they will finally automate half of all office work and take all of those office paychecks as revenue.
>>1023641This one's completely trash. The chinese one is way better. But really who cares if it cant' retopo it. Where the fuck am I gonna use a 1.5 mil sculpt.
>>1024134alright their studio thing can retopo & segment even. if you try to retopo the whole model it sucks but if you do it piece by piece it's close to a usableish result. their retopo model is not in the public and is by far the most intensive takes a relatively long ass time compared to the rest of the process. I have no clue how you'd combine at all desu. another AI model to fill in the gaps? it can already segment.
>>1023844This is Z-image Turbo today. It generate that in minutes.I can't imagine what Wan will be tomorrow.
>>1023641It's over for blender porn animators unless they switch to animating with AI. I can just take this pinup render of Mother Miranda by word2 and animate it like this. I can also just animate in game screenshots, for max authenticity
>>1024221I just figured out about 20 steps Ksampling.But cfg values not so much.We can discuss this on /ldg/
https://files.catbox.moe/4qfii9.mp4It's over for animation studiosThis is from Wan2.2 Infinite loop that dropped today.
>>1024241still looks like your typical ai slop, not sure what progress you want to point out.
>>1024246You don't get it, it's not about the subject, just an example, now you can generate more than a 5 seconds clip before details deteriorates. They implemented self checks and stuff.You just have to write several prompts and time them.
>>1024247the detail deteriorates literally at the first generated frame
>>1024247You absolute mongoloid.
