Trained on the same 69 images as v2.0, but with 12 vectors per token instead of 2, learning rate of 0.0015. Best results obtained at 18,000 steps.
I'm very happy with this - it needs far less 'steering' with prompts. Almost all pics I generated were recognisably her, about 1 in 5 are spot on.
Tests produced good results in most realistic models (Ares is the exception), with superb results in Cyberrealistic, Deliberate, DeliberateRealistic and RealisticVision. My preference is for CFG 10 with this TI, YMMV.
Trained on the same 69 images as v2.0, but with 12 vectors per token instead of 2, learning rate of 0.0015. Best results obtained at 18,000 steps.
I'm very happy with this - it needs far less 'steering' with prompts. Almost all pics I generated were recognisably her, about 1 in 5 are spot on.
Tests produced good results in most realistic models (Ares is the exception), with superb results in Cyberrealistic, Deliberate, DeliberateRealistic and RealisticVision. My preference is for CFG 10 with this TI, YMMV.