thanks I don't do a good job of advertising it but it likely has one of the largest dataset 850 images to be exact, it required a stupid amount of training steps to accommodate. that's why it was trained up to 101,000 steps, my only regret is I should have included more closeups in the set. Thats why I think it works so well with cyber realistics turbo checkpoint specifically the V5 version. It seems to fill the gaps in my lora.

@DocHector Its the exact same Lora because lora's trained on Zimage base work with the turbo model , I created this entry so that people would realize that. Lora's trained on the turbo de-distilled tool don't work with base though and are actually less effective turbo loras.

@WhiteRider yeah, its because my dataset was 850 images. even though 85000 put me at 100 epochs I let it run for 142000 steps total, it seemed to work best on 101,000. However the base model I just released is actual merged with my 130000 step version. It might be a little overcooked hahah. But it still gives good results. I might release a version with only the 125,000 merge.

@WhiteRider Training a lora for Zimage seems to need at least 115 epochs so if you have 500 images in your data set you need to train for 57,000 steps min but allow it to run over a few thousand if you can. overtraining my lora was super useful for testing to see how it reacted. rent a server do it on a 5090 be prepared to spend the time, my bean effect training ran for days because I chose a cheap slow GPU and in the end was a huge mistake. But because I had already let it run for days it became sunk cost issue that cost me about 45 dollars in server time. Never again! always choose the 5090! also even if you want to use your lora on turbo train it on Base! Unless some of these new Turbo models have strayed very far from the base model that will yield the best results. Any lora train on the true base model will be compatible with base and Turbo. Mine work that way. Also the bean effect lora works fine on cyberrealistic turbo and others I've tried.