Implementation of EasyControl

EasyControl: Adding Efficient and Flexible Control for Diffusion Transformer

Yuxuan Zhang, Yirui Yuan, Yiren Song, Haofan Wang, Jiaming Liu
Tiamat AI, ShanghaiTech University, National University of Singapore, Liblib AI

Features

• Motivation: The architecture of diffusion models is transitioning from Unet-based to DiT (Diffusion Transformer). However, the DiT ecosystem lacks mature plugin support and faces challenges such as efficiency bottlenecks, conflicts in multi-condition coordination, and insufficient model adaptability, particularly in zero-shot multi-condition combination scenarios where these issues are most pronounced.
• Contribution: We propose EasyControl, an efficient and flexible unified conditional DiT framework. By incorporating a lightweight Condition Injection LoRA module, a Position-Aware Training Paradigm, and a combination of Causal Attention mechanisms with KV Cache technology, we significantly enhance model compatibility, generation flexibility, and inference efficiency.

Download

You can download the model directly from Hugging Face. Or download using Python script:

from huggingface_hub import hf_hub_download
hf_hub_download(repo_id="Xiaojiu-Z/EasyControl", filename="models/canny.safetensors", local_dir="./models")
hf_hub_download(repo_id="Xiaojiu-Z/EasyControl", filename="models/depth.safetensors", local_dir="./models")
hf_hub_download(repo_id="Xiaojiu-Z/EasyControl", filename="models/hedsketch.safetensors", local_dir="./models")
hf_hub_download(repo_id="Xiaojiu-Z/EasyControl", filename="models/inpainting.safetensors", local_dir="./models")
hf_hub_download(repo_id="Xiaojiu-Z/EasyControl", filename="models/pose.safetensors", local_dir="./models")
hf_hub_download(repo_id="Xiaojiu-Z/EasyControl", filename="models/seg.safetensors", local_dir="./models")
hf_hub_download(repo_id="Xiaojiu-Z/EasyControl", filename="models/subject.safetensors", local_dir="./models")

If you cannot access Hugging Face, you can use hf-mirror to download the models:

export HF_ENDPOINT=https://hf-mirror.com
huggingface-cli download --resume-download Xiaojiu-Z/EasyControl --local-dir checkpoints --local-dir-use-symlinks False

Usage

Here's a basic example of using EasyControl. For more details, please follow the instructions in our GitHub repository:

Model Initialization

import torch
from PIL import Image
from src.pipeline import FluxPipeline
from src.transformer_flux import FluxTransformer2DModel
from src.lora_helper import set_single_lora, set_multi_lora

def clear_cache(transformer):
    for name, attn_processor in transformer.attn_processors.items():
        attn_processor.bank_kv.clear()

# Initialize model
device = "cuda"
base_path = "FLUX.1-dev"  # Path to your base model
pipe = FluxPipeline.from_pretrained(base_path, torch_dtype=torch.bfloat16, device=device)
transformer = FluxTransformer2DModel.from_pretrained(
    base_path, 
    subfolder="transformer",
    torch_dtype=torch.bfloat16, 
    device=device
)
pipe.transformer = transformer
pipe.to(device)

# Load control models
lora_path = "./models"
control_models = {
    "canny": f"{lora_path}/canny.safetensors",
    "depth": f"{lora_path}/depth.safetensors",
    "hedsketch": f"{lora_path}/hedsketch.safetensors",
    "pose": f"{lora_path}/pose.safetensors",
    "seg": f"{lora_path}/seg.safetensors",
    "inpainting": f"{lora_path}/inpainting.safetensors",
    "subject": f"{lora_path}/subject.safetensors",
}

Single Condition Control

# Single spatial condition control example
path = control_models["canny"]
set_single_lora(pipe.transformer, path, lora_weights=[1], cond_size=512)

# Generate image
prompt = "A nice car on the beach"
spatial_image = "./test_imgs/canny.png"

image = pipe(
    prompt,
    height=720,
    width=992,
    guidance_scale=3.5,
    num_inference_steps=25,
    max_sequence_length=512,
    generator=torch.Generator("cpu").manual_seed(5),
    spatial_images=[spatial_image],
    cond_size=512,
).images[0]

# Clear cache after generation
clear_cache(pipe.transformer)

Multi-Condition Control

# Multi-condition control example
paths = [control_models["subject"], control_models["inpainting"]]
set_multi_lora(pipe.transformer, paths, lora_weights=[[1], [1]], cond_size=512)

prompt = "A SKS on the car"
subject_images = ["./test_imgs/subject_1.png"]
spatial_images = ["./test_imgs/inpainting.png"]

image = pipe(
    prompt,
    height=1024,
    width=1024,
    guidance_scale=3.5,
    num_inference_steps=25,
    max_sequence_length=512,
    generator=torch.Generator("cpu").manual_seed(42),
    subject_images=subject_images,
    spatial_images=spatial_images,
    cond_size=512,
).images[0]

# Clear cache after generation
clear_cache(pipe.transformer)

Usage Tips

• Clear cache after each generation using clear_cache(pipe.transformer)
• For optimal performance:
  • Start with guidance_scale=3.5 and adjust based on results
  • Use num_inference_steps=25 for a good balance of quality and speed
• When using set_multi_lora api, make sure the subject lora path(subject) is before the spatial lora path(canny, depth, hedsketch, etc.).

Disclaimer

The code of EasyControl is released under Apache License for both academic and commercial usage. Our released checkpoints are for research purposes only. Users are granted the freedom to create images using this tool, but they are obligated to comply with local laws and utilize it responsibly. The developers will not assume any responsibility for potential misuse by users.

Citation

@misc{zhang2025easycontroladdingefficientflexible,
      title={EasyControl: Adding Efficient and Flexible Control for Diffusion Transformer}, 
      author={Yuxuan Zhang and Yirui Yuan and Yiren Song and Haofan Wang and Jiaming Liu},
      year={2025},
      eprint={2503.07027},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2503.07027}, 
}