Cuties, it's close to the anniversary of when I first started seriously working with AI, so I wanted to make my first LoRA! ✺◟( ˃ᴗ˂ )◞✺ aww yiss!

💨 This ComfyUI Workflow is for Z-Image Turbo!

Download the LoRAs here: https://civarchive.com/models/2253443/sues-z-body-bikini-beach-by-freyja-pixel

📸 768x1280 or similar portrait is the recommended orientation and latent image size!

I used Musubi Tuner to make a Z-Image LoRA using some of my first ever AI generations from early 2025! The 32 training images were made in January 2025 using Wai-Illustrious version 8 and Automatic1111, and the LoRA was made in December 2025! I used my customized Chat GPT 5.2 to write the base captions, and then I engineered and iterated upon them with Z-Image and the alpha-test LoRA to fine tune the training! I also increased the steps up to 480, and made 2 different versions using different VAE training! The AE version uses the base Z-Image ae.safetensor for the training, and the UltraFLUX uses a community modded vision version of the Flux VAE. I then created a customized workflow based on my observations during image testing!

⚠️⚠️⚠️ KNOWN ISSUES:

These LoRAs were trained off of images that do not have good hands and feet! I was still learning when I genned the images, and the detailed hands and feet are few and far between! I did my best, but you will likely have to roll a few gens and seed combinations! You could also try prompting more, but I have not tried that!

😇 THIS LORA IS SFW. It does not include any nudity in it's training images!

🛠️ Everything in this workflow has been custom tested based on my experience and community recommendations! I recommend keeping most of the settings as-is! If you want to tinker, think about changing the Ksampler to a completely randomized seed set up, or change the scheduling to "res_2" and "simple" which are the Z-Image defaults. Finally, it finishes off with a 1 percent noise buffer for cinematic quality!

⏱️ My gen times were around 55-65 seconds with full bf16 weights and one pass! For the full set up on my 12 GB VRAM, it was about 100 seconds for both passes! For the Q-8 GGUF, I had 100 - 120 seconds!

💪 The LoRAs themselves work excellently between .4 - .6 weight, though I found my best results at between .48 - .52, favoring the later .52 if I'm running just one pass/single LoRA.

👙 The LoRAs are great with skin detail, anatomy, perspective, and bikinis! They are trained specifically on "an anime illustration of a blonde Caucasian woman at a beach at sunset", and I haven't really tested much outside of that subject and scene. I loved seeing Sue wearing white bikinis at the beach, so that was one of the first image sets I made!

🖥️ The workflow uses both LoRAs in a dual-pass style, latent hi-res 2x upscaling, with the higher quality Flux version on the refining second pass! If you like, you can bypass most everything that is in a colored box and run the Flux LoRA version by itself for base res images! This will result in lightning fast generations! Anywhere from 15 - 30 seconds on my set up!

Color Code:

💙💙💙 Power LoRA loader with AE LoRA, Single Model-based LoRA loader with Flux LoRA, Second pass K-Sampler nodes

💜💜💜 Hi-res latent upscale, noise buffer