From my brief testing it does get very close to how MarigoldXL looks but due to it being on Flux doesn't look fully there, which is to be expected. V2 is made using a more balanced dataset of 45 images. It's been trained on Schnell but works well on both Dev and Schnell, if using Dev I'd recommend to run it with the hyper-16 lora at 18 steps. ---- V2 is made using a more balanced dataset of 45 images.