LibreFLUX-ControlNet

Update - 4/10/2026

• Retrained this model on laion2b-squareish-1536px
• I tripled the control layers, to get better guidance

Fun Facts

• Trained exclusively on images generated by Segment Anything (SAM)
• Uses SAM style images as input, outputs photorealistic images
• Trained at 1024x1024 resolution, inference works best at 1.5k and up
• Trained on 320K segmented images from laion2b-squareish-1536px
• Base model is LibreFLUX ( de-distilled FLUX )

Showcases

Extra Details

• I built this repo to train the model: https://github.com/NeuralVFX/LibreFLUX-ControlNet
• Trained in same non-distilled fashion as LibreFLUX
• Uses Attention Masking
• Uses CFG during Inference ( allows negative prompting )
• Inference code roughly adapted from: https://github.com/bghira/SimpleTuner

ComfyUI

• I've made some custom nodes for this: https://github.com/NeuralVFX/LibreFLUX-ComfyUI

Compatibility

pip install -U diffusers==0.32.0
pip install -U "transformers @ git+https://github.com/huggingface/transformers@e15687fffe5c9d20598a19aeab721ae0a7580f8a"

Low VRAM:

pip install optimum-quanto

Load Pipeline

import torch
from diffusers import DiffusionPipeline

model_id = "neuralvfx/LibreFlux-ControlNet"  
device = "cuda" if torch.cuda.is_available() else "cpu"

dtype  = torch.bfloat16 if device == "cuda" else torch.float32

pipe = DiffusionPipeline.from_pretrained(
    model_id,
    custom_pipeline=model_id,
    trust_remote_code=True,   
    torch_dtype=dtype,
    safety_checker=None        
).to(device)

Inference

from PIL import Image
from torchvision.transforms import ToTensor

# Load Control Image
cond = Image.open("examples/libre_flux_control_image.png")
cond = cond.resize((1024, 1024))

# Convert PIL image to tensor and move to device with correct dtype
cond_tensor = ToTensor()(cond)[:3,:,:].to(pipe.device, dtype=pipe.dtype).unsqueeze(0)

out = pipe(
  prompt="many pieces of drift wood spelling libre flux sitting casting shadow on the lumpy sandy beach with foot prints all over it",
            negative_prompt="blurry",
            control_image=cond_tensor,  # Use the tensor here
            num_inference_steps=75,
            guidance_scale=4.0,
            height =1024,
            width=1024,
            controlnet_conditioning_scale=1.0,
            num_images_per_prompt=1,
            control_mode=None,
            generator= torch.Generator().manual_seed(32),
            return_dict=True,
        )
out.images[0]

Load Pipeline ( Low VRAM )

import torch
from diffusers import DiffusionPipeline
from optimum.quanto import freeze, quantize, qint8

model_id = "neuralvfx/LibreFlux-ControlNet" 
device = "cuda" if torch.cuda.is_available() else "cpu"
dtype  = torch.bfloat16 if device == "cuda" else torch.float32

pipe = DiffusionPipeline.from_pretrained(
    model_id,
    custom_pipeline=model_id,
    trust_remote_code=True,    
    torch_dtype=dtype,
    safety_checker=None         
)

quantize(
    pipe.transformer,
    weights=qint8,
    exclude=[
        "*.norm", "*.norm1", "*.norm2", "*.norm2_context",
        "proj_out", "x_embedder", "norm_out", "context_embedder",
    ],
)

quantize(
    pipe.controlnet,
    weights=qint8,
    exclude=[
        "*.norm", "*.norm1", "*.norm2", "*.norm2_context",
        "proj_out", "x_embedder", "norm_out", "context_embedder",
    ],
)
freeze(pipe.transformer)
freeze(pipe.controlnet)

pipe.enable_model_cpu_offload()