Hitomi Tanaka for Flux

The best Hitomi Tanaka Lora ever made. I do not dabble in mediocrity. As an homage, I gave her 36P cups that generated glorious results.

Civit is so aggressive that I am forced to post photos of her in sweaters, even a NORMAL blouse is way too proactivity, but trust me when I say the actual results are INCREDIBLE

I looked at the existing models here and they are honestly pretty bad. Low-quality results and clearly not much attention to detail when making the Lora. Pretty shocking there were no high-quality loras of this ultra-famous model. That has now been fixed.

To make this five stars I used half and half-clothed samples. 9 high-res photos of her face from different angles. Photoshopped in some pubic hair and to avoid any blurring. I extended her just a touch so it was roughly 1.5 inches past her belly button. This provides more consistent generations. The dataset is available by request.

PSA: Hitomi's weighed 7.4 pounds in a video, here is the image. That weight, which appears accurate, makes her a 34K cup. During her last billings, she was listed as a Japanese O cup, which is a US N cup or UK JJ cup. That would place her as 36N, which is grossly incorrect as her would need to be 13 pounds (combined) to fit that . The photo I just shared shows her at 7.4 pounds, there is not much difference between that and when she retired, which means she clearly did not put on DOUBLE the weight. So the listing is heavily exaggerated in order to sell more content. I properly measured tomorrow, and she would likely be a 36J cup. At any rate, I intentionally adjusted her photos in the dataset to more properly show her at 36N. You will be pleasantly surprised with the results.

Training details:

• Steps: 5,000, 2 batch (ai-toolkit)

• Samples: 50 samples per image

• Trigger word: H1t0m1T4naka

• Linear/Alpha 32

• Learning rate: 0.0001 (bf16)

• Image size: 1024x1024

• Dataset: 48 detailed photos of her face, head, and body

• 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.