flux-uncensored-nf4

Summary

Flux base model merged with uncensored LoRA, quantized to NF4. This model is not for those looking for "safe" or watered-down outputs. It’s optimized for real-world use with fewer constraints and lower VRAM requirements, thanks to NF4 quantization.

Specs

Model: Flux base
LoRA: Uncensored version, merged directly
Quantization: NF4 format for speed and VRAM efficiency

Usage

Not so much for plug-and-play model, but pretty straight forward (script from sayak [https://github.com/huggingface/diffusers/issues/9165#issue-2462431761])

Please install pip install -U bitsandbytes to proceed.

"""
Some bits are from https://github.com/huggingface/transformers/blob/main/src/transformers/modeling_utils.py
"""

from huggingface_hub import hf_hub_download
from accelerate.utils import set_module_tensor_to_device, compute_module_sizes
from accelerate import init_empty_weights
from convert_nf4_flux import _replace_with_bnb_linear, create_quantized_param, check_quantized_param
from diffusers import FluxTransformer2DModel, FluxPipeline
import safetensors.torch
import gc
import torch

dtype = torch.bfloat16
is_torch_e4m3fn_available = hasattr(torch, "float8_e4m3fn")
ckpt_path = hf_hub_download("shauray/flux.1-dev-uncensored-nf4", filename="diffusion_pytorch_model.safetensors")
original_state_dict = safetensors.torch.load_file(ckpt_path)

with init_empty_weights():
    config = FluxTransformer2DModel.load_config("shauray/flux.1-dev-uncensored-nf4")
    model = FluxTransformer2DModel.from_config(config).to(dtype)
    expected_state_dict_keys = list(model.state_dict().keys())

_replace_with_bnb_linear(model, "nf4")

for param_name, param in original_state_dict.items():
    if param_name not in expected_state_dict_keys:
        continue
    
    is_param_float8_e4m3fn = is_torch_e4m3fn_available and param.dtype == torch.float8_e4m3fn
    if torch.is_floating_point(param) and not is_param_float8_e4m3fn:
        param = param.to(dtype)
    
    if not check_quantized_param(model, param_name):
        set_module_tensor_to_device(model, param_name, device=0, value=param)
    else:
        create_quantized_param(
            model, param, param_name, target_device=0, state_dict=original_state_dict, pre_quantized=True
        )

del original_state_dict
gc.collect()

print(compute_module_sizes(model)[""] / 1024 / 1204)

pipe = FluxPipeline.from_pretrained("black-forest-labs/flux.1-dev", transformer=model, torch_dtype=dtype)
pipe.enable_model_cpu_offload()

prompt = "A mystic cat with a sign that says hello world!"
image = pipe(prompt, guidance_scale=3.5, num_inference_steps=50, generator=torch.manual_seed(0)).images[0]
image.save("flux-nf4-dev-loaded.png")

this README has what you'd need, it's a merge from Uncensored LoRA on CivitAI