There is nothing I hate more than censorship in text-to-image models. It's strange that sexualized depictions of teenagers are allowed, but a classical figure like Venus without clothes is censored. Who decides what's acceptable in AI models? If I had the resources, I’d train my own model free from these restrictions. Censored models like Flux.1 distort female anatomy. Models like Kontext and Flux.2 are worse. LoRAs on HuggingFace that can properly present the female body got banned. That’s why I decided to build my own NSFW QLoRA. I had to manually create prompts for each image because VLM models are too censored to help. I believe one day, we’ll be able to train our own models on consumer hardware without chains from these regulations.

The QLoRA fine-tuning process of pachu_torres_style_lora_flux_nf4 takes inspiration from this post (https://huggingface.co/blog/flux-qlora). The training was executed on a local computer with 1500 timesteps and the same parameters as the link mentioned above, which took around 10 hours on 8GB VRAM 4060. The peak VRAM usage was around 7.7GB. To avoid running low on VRAM, both transformers and text_encoder were quantized. All the images generated here are using the below parameters

Height: 512
Width: 512
Guidance scale: 5
Num inference steps: 20
Max sequence length: 512
Seed: 0

Usage

python

import torch
from diffusers import FluxPipeline, FluxTransformer2DModel
from transformers import T5EncoderModel

text_encoder_4bit = T5EncoderModel.from_pretrained(
    "hf-internal-testing/flux.1-dev-nf4-pkg", subfolder="text_encoder_2",torch_dtype=torch.float16,)

transformer_4bit = FluxTransformer2DModel.from_pretrained(
        "hf-internal-testing/flux.1-dev-nf4-pkg", subfolder="transformer",torch_dtype=torch.float16,)

pipe = FluxPipeline.from_pretrained("black-forest-labs/FLUX.1-dev", torch_dtype=torch.float16,
                                    transformer=transformer_4bit,text_encoder_2=text_encoder_4bit)

pipe.load_lora_weights("je-suis-tm/pachu_torres_style_lora_flux_nf4",
                       weight_name='pytorch_lora_weights.safetensors')

prompt="Pachu Torres style, this image depicts a dark and mysterious scene featuring two women. The woman in the back, masked with a black bunny mask (eyes visible), is blindfolding a woman in the front. The blindfolded woman, in a black dress, clings to the masked woman's arm. She wears a black choker with metal leash attached. A striking blue background, framed by a darker border, contrasts with the foreground's shadowy tones. The image evokes a power dynamic and leaves the viewer intrigued by the domination and submission"

image = pipe(
            prompt,
            height=512,
            width=512,
            guidance_scale=5,
            num_inference_steps=20,
            max_sequence_length=512,
            generator=torch.Generator("cpu").manual_seed(0),            
        ).images[0]

image.save("pachu_torres_style_lora_flux_nf4.png")

Trigger words

You should use Pachu Torres Style to trigger the image generation.

Download model

Download them in the Files & versions tab.