I've had good generation consistency with Auto/Kerras/Polyexpo on the SDEs and Euler A at LoRA: 0.8 / trigger: 1.1. When it comes to Clip, both 1 and 2 produce good results.
SD: Seems to be consistent across "zoom levels"
SDXL: Definitely prefers portrait or full torso shots.
Pony: The first one I've tried in a while, it seems to behave fairly well across checkpoints, including my experimental Equus model.
Sarah Catherine Hook
Born April 21, 1995 (29 years)
Sarah Catherine Hook is an American actress and singer. Born and raised in Montgomery, AL, she went on to SUNY Purchase College to pursue a degree in Vocal Performance: Opera where she received her Bachelor of Music. During her time at Purchase, she began to pique her interest in acting and attended the Atlantic Theater Company Summer Intensive. Since then, she has continued to take both acting and singing courses around New York. Sarah Catherine is known for The Conjuring 3, Law & Order: SVU, and Living in Harmony Sodastream.
Description
This comes as a result of experimentation with FluxGym, which is pretty painless to get up and running, and seems to be a great way to train Flux LoRAs on your local machine. The downsides? Well, I probably just melted an arctic shelf not only with the amount of energy used to, but also the heat produced by, the GPU. Well, that mainly melted me... The other downside is that it took 10.5 hours on a 4060 Ti/16GB.
Anyway, this was created using the same dataset as the other 3 models.
The example images were generated on several models using a smattering of Euler/DEIS and Normal/KL Optimal/SGM Uniform:
my_stinky_model (named in frustration when attempting Flux checkpoint merges)
flux1-dev-Q4_K_S.gguf
flux1DevFlux1Schnell_devNF4UNET
Fluxzilla
FluxFusionDS_v0_Q4_K_S