Absolutely MINDBLOWING Lora of JAV actress Rika Fujishita!! I honestly cannot conceive how this Lora could be any better. It's better than real life. I have truly outdone myself with this Lora. The results are mouthwatering. Even basic prompting produces amazing results.
V2 has an improved chest with monster hooters past her stomach. Details in her face and bush are better in V2. Example. If you prefer her more natural size use V1.
First, I created a proper Lora with her accurate bra size. The result was pretty good, but Flux absolutely HATES nudity and kept trying to make her chest smaller. So, I took some of the best generations and tossed that BACK into my dataset. Then carefully photoshopped all 17 images of her chest to give her PROPER belly floppers. That way Flux has two choices: Show her adjusted breast size or make them smaller, which would turn out to be accurate to real life.
Try some prompts like bending over. Side view. Or lying down. Lots of options. Prompt "She is wearing a business suit and ripping her shirt open" like this masterpiece, or this.
The actress is 34G cup in real life but I made her with 34L cups, which better suits her frame.
In my opinion, she is gorgeous and has many great videos. Here are two BEST movies she has ever made. First (great throughout) and Second (the last scene is EPIC, let it play through)
Training details:
Steps: 3,000, 2 batch (ai-toolkit)
Samples: 50 samples per image
Trigger word: R1k4Fuj1sh1t4 (not required)
Linear/Alpha 32
Learning rate: 0.0001 (bf16)
Image size: 1024x1024
Dataset: 29 detailed photos. 9 of her face. 18 of her nude body in various positions. 3 in outfits.
Training prompts: Prompted out makeup and jewelry
ComfyUI usage:
Please note that all of my sample images double as a ComfyUI workflow! Simply drag any of the sample images into ComfyUI and it will open the entire workflow, set the seed, and use my same configuration. I typically use LLM generation but chose not to do so for these tests; instead, I used a more basic workflow.
Comfy settings:
Steps: 35
Guidance: 3.5
Sampler: DPM++ 2M, Beta
You can get good results with Euler but DPM++2M produces the BEST results. Now if you stack a few loras for generation, then Euler might be your best bet since those other loras were LIKELY trained on Euler.