Modelpublic

olob0/z-image-turbo-brazilian-male-realism-lora-v2

sourceHugging Facecc-by-4.0updated 3mo agoView on Hugging Face

2likes107downloads

Model Card

Brazilian Male Realism LoRA V2 (Z Image Turbo)

This is my second attempt to train a LoRA for Z Image Turbo, fixing some of the issues I had with the first one. This one is focused on improving the realism of Brazilian/Latin American pics, keeping the same quality of the base model.

The goal is to reduce the the "realistic AI look"/"cozy-perfect" of the highly realistic images generated by Z Image Turbo, resulting in more natural proportions and imperfections, typical of real people, without exaggeration. In other words, it attempts to reduce the super-detailed, perfect appearance, making the images more similar to those taken with a cell phone.

Trained with 100+ customized images of Brazilian men, with much more variations in skin tone and age compared to the V1. It was trained with more AI-generated images, using ChatGPT-Image-2.0 for generations and Google Nano Banana 2 for upscaling some images in the original dataset.

<div align="center"> <p>Comparisons:</p>

<p><i>Open in full screen, will loop</i></p>

<p>As you can see, some images may work better with low intensity, and others with high intensity.</p> </div>

Download model

You can Download them in the Files & versions tab. Named z-image-turbo-brazilian-male-realism-lora-v2-bf16.safetensors

Usage

—Recommended LoRA strength: It varies according to your preference. The closer to 0.5, the more similar the result is to the base model. The closer to 1.0, the more LoRA will interfere with the generation. Some images may work better with low intensity, and others with high intensity. Do your tests! Try something between 0.6 and 0.8 (~0.75).

—Guidance: (CFG Scale): It is the same as the base model, 1.0.

—Steps: 8-10. It is the same as the base model, 8.

—Trigger word: BRZMLRELSM.

<div align="center"> <p>Same seed, same prompt, different LoRA strength:</p>

<p>As you can see, some images may work better with low intensity, and others with high intensity.</p> </div>

How to prompt

For best results, I recommend trying to describe each element of the prompt semantically and in detail (You can use ChatGPT, Gemini, Grok, Claude, or any other AI of your choice to "hyper-describe" your prompts). I recommend using "piggybacking," that is, using the knowledge of the base model to your advantage. Try including terms like "Brazilian man" or "Latino man" along with the triggerword. Take a look at the examples I added!

Training details

—Base Model: Tongyi-MAI/Z-Image-Turbo
—LoRA Network: Rank 16 (Linear) / Rank 8 (Conv)
—LoRA Alpha: 16 (Linear) / 4 (Conv)
—Optimizer: AdamW
—Learning Rate: 1.5e-05
—Training Steps: 1600
—Batch Size: 1
—Resolutions: 512, 768, 1024

Limitations

—Sometimes, it can distort/corrupt text in images, especially small text.
—For some reason, certain internal patterns, such as interior walls made of large blocks, may become slightly distorted if there isn't a specific description of them in the prompt. In other words, always describe what the background of the image will look like.

Examples

\ All prompts are generated and enhanced using `Gemini 3.1 Pro`*.

Negative prompt used for all images:

smooth skin, silky skin, professional, studio lighting, airbrushed, smooth skin, plastic, fake, cgi, render, 3d, makeup, lipstick, feminine, woman, (low quality, worst quality, lowres, blurry, jpeg artifacts, noise, grain, distorted, deformed, bad anatomy, bad proportions, asymmetry, ugly, uncanny, creepy, asian, east asian, southeast asian, chinese, japanese, korean, thai, vietnamese, asian face, asian features, monolid eyes, epicanthic fold, flat nose bridge, wide cheekbones, anime face, k-pop face, sad, depressed, crying, gloomy, idol look, muscles)

License

This model is licensed under CC-BY-4.0.

You are free to use, modify and distribute this LoRA, including for commercial purposes, as long as proper credit is given to the original author.

Acknowledgements

—Z Image Turbo base model by Tongyi-MAI
—Training adapter by ostris

Citation

If you use this LoRA in your work, please cite:

bibtex

@misc{lobo_brazilian_male_realism_lora_v2,
  title = {Brazilian Male Realism LoRA for Z Image Turbo v2},
  author = {Gabriel Lobo},
  year = {2026},
  howpublished = {url{https://huggingface.co/olob0/z-image-turbo-brazilian-male-realism-lora-v2}},
}