Hello everyone. I am Dr. Furkan Gözükara. PhD Computer Engineer. SECourses is a dedicated YouTube channel for the following topics : Tech, AI, News, Science, Robotics, Singularity, ComfyUI, SwarmUI, ML, Artificial Intelligence, Humanoid Robots, Wan 2.2, FLUX, Krea, Qwen Image, VLMs, Stable Diffusion
I repost also here... maybe it's the correct section: Hi Guys... it's black-friday/cybermonday time... do you have any advice on what to buy to make flux/sdxl/pdxl models locally? I mean: what minimum requirements for a graphic card that can handle those kind of jobs... something cheap, as I do training only to understand, not for selling my works around, so I can bear also longest trainings... woud appreciate model cards or laptops adequately equipped advices, thanks to all!
I went for cheapest CPU cheapest PC with 4070 Ti Super 16 GB and it's holding up. I'd go for 64 GB RAM now though because the latest offloading techniques by kohya and OneTrainer need lots of RAM. everything can change though by tomorrow when a new model comes out. but for the current models this is a very cost-effective setup
btw @Furkan Gözükara SECourses since you've complained about the OneTrainer UI on massed compute again: the cloud extension is very mature by now. a few people use it. you can run OneTrainer locally but train remotely. you might want to try it. it works best on runpod, but also works with any other cloud that gives SSH access
during the grid test in swarm, is it possible to get full quality images? I'm getting jpg ones, I even changed to png in the settings. It's like during grid, the pic quality is less good.
How do you determine which model or LoRA is better using a GRID? I conducted a test between 10 fine-tuned models, and on all of them, the similarity to a real person looks excellent! What do you rely on? Perhaps you use some special prompts to choose one model?
The thing is, they all look like my character, and it's quite difficult to choose. The question is more about the fact that next time I can use fewer epochs for training, as I understand.