This repository contains the official LoRA adapter for black-forest-labs/FLUX.1-dev fine-tuned using CMO (Correlation-Weighted Multi-Reward Optimization), our ECCV 2026 work on improving compositional generation capabilities. This variant was trained with the CMO reward together with the HPSv2 reward for human preference alignment.

🚀 Usage

Below is the code to load and merge the LoRA adapter with the base FLUX.1-dev model.

python

import torch
from diffusers import FluxPipeline
from peft import PeftModel

model_id = "black-forest-labs/FLUX.1-dev"
lora_ckpt_path = "Bruece/FLUX.1-dev-CMO-HPSv2"
device = "cuda"

pipe = FluxPipeline.from_pretrained(model_id, torch_dtype=torch.bfloat16)
pipe.transformer = PeftModel.from_pretrained(pipe.transformer, lora_ckpt_path)
pipe.transformer = pipe.transformer.merge_and_unload()
pipe = pipe.to(device)

prompt = "a photo of a black kite and a green bear"
image = pipe(
    prompt,
    height=512,
    width=512,
    num_inference_steps=40,
    guidance_scale=4.5
).images[0]

image.save("flux_cmo_hpsv2_lora.png")

🖼️ Qualitative Results

ConceptMix (Link)

GenEval 2 (Link)

T2I-CompBench (Link)

🛠️ Training Details

Base Model: FLUX.1-dev
Algorithm: Correlation-Weighted Multi-Reward Optimization (CMO)
Additional Reward: HPSv2 human preference reward, trained jointly with the CMO reward
Precision: bfloat16

📜 Citation

If you find this ECCV 2026 model useful for your research, please cite:

bibtex

@article{wi2026correlation,
  title={Correlation-Weighted Multi-Reward Optimization for Compositional Generation},
  author={Wi, Jungmyung and Kim, Hyunsoo and Kim, Donghyun},
  journal={arXiv preprint arXiv:2603.18528},
  year={2026}
}