/g/ - Technology

>>780 (OP) 
HunyuanWorld1.0, a new 3d-world generating model, seems great, but I was unable to make it run on a 4090 and 64 gigs of ram, seems you need about 100 or so, even with my script editing to make it more efficient.

The part it can't handle is the generation, the Panorama and world masks were surprisingly not very intensive, I could not run it originally either, but switching the model to use SDXL over FLUX, made it take shockingly little compute resources, could probably run on a 12gb model. 

The actual mesh generation itself is poorly optimized by the gooks and has a ton of RAM eating processes running in parallel instead of sequentially, I don't have the time and it is not worth it for me to fix such a huge flaw.

The panorama however can be used with Hunyuan3D Mesh Gen model locally to create the worlds however, but it's more intensive process obviously, good for HDRI generation though.

Hunyuan3d2.1 is amazing but the topology is mid, only real downside, the mesh generation takes little resources, texturing is more, but I could "max" out all settings on texture gen with my system, just high requirements for the model itself to load, after that increasing settings is gentle on compute.

OneTrainer is the best LoRA trainer I am aware of, works well, Hydrus for image tagging, the documentation for Hydrus is pretty bad, and UI is horrendous, but working with it is really smooth and great.

AI only knows what you know. It intellectually does not go much further than that. In the unpaid models, I've only found chat GPT with the reasoning version to have any decent hitting power. But it thinks things which aren't real because it doesn't actually understand the subject matter, so you have to guide it's reasoning.

Only thing these things are good for is showing you what you missed in your own blind spots. Like a brainstorming program which you gently guide using your understanding of what's realistic. For example, if you ask it to simulate a battle using weights and numbers, it usually can't. It will have a series of improvisations to fool the user into thinking it can, but actually the basis for those ideas are all hokum. It's like that with everything. Some of the chatbots you have to basically guide by the hand to get them to do the things you want, they are that passive and dependent on the user's framing and input.

AI is a scam. Now we don't have a term anymore to name an actual AI with but need to invent a new one, like "real AI", which is about as retarded as the time we live in.

Also: Isn't it funny how they spent so much time name-calling and alienating the people with pattern recognition and enough of a spine to call a spade a spade only to then invent pattern recognition systems that turned out to do things they didn't want them to do but are forced to advertise under a wrong label because that's all they have going now ?

>>783
There's a lot of hardcore shilling for it too. They are paying for favorable and hype building media coverage in many of these newspapers. Many of these planted pseudo-viral stories about them doing amazing things and all the rest. Mechanical turking them for show I suspect. LLM is not AI in the real sense