*UPDATE* I am including a variety of variants of this model rather than posting them as updates. There is a number attached to each version that represents how strongly the training data I added impacts the Base model, higher numbers are stronger in terms of bean effect. highest is 130 and steps down from there, the numbers are representing training steps in thousands, 130 is 130,000 steps of training on the 850 image dataset. you would think this would overtrain or cook the model but it turns out if you have enough variety in the dataset it doesn't. Zimage is a hungry model I think its been neglected because people don't want to put in the effort for good training. personally I think 125 is the sweet spot but 115 is better for creativity. 130 is probs too much.

This is a Q8 Merge of my Bean effect lora into Zimage Base then quantized Q8

essentially this is a very functional Zimage Base model that does NSFW very very well.

I don't know about anyone else but I still use the Base model in my workflows to generate the initial image because zimage turbo isn't great at variety. So then I use Turbo to refine the image............ However all example images are generated only with this model and the z image fun distill lora to reduce generation steps. Its actually pretty good flying solo seems to produce better more photorealistic images than base. I recommend using the zimage fun distill lora at 0.75 and Euler with beta scheduler at 16-20 steps

there are some weak spots, blowjobs defiantly, there are a few images in the training data but its a weak point due to my tastes. as states the bean effect is geared towards shapely booty in all poses, with sexual positions baked in at same power. but they lean towards views from behind but not totally at all. There's a lot of focus on real women with large and small breasts, but because the training was loaded towards a variety of sexual and solo booty images you will find it does it better than any other.