FLUX.1-dev-CMO-HPSv2

FLUX.1-dev-CMO-HPSv2

<p align="center"> <img src="./cmo_logo_tight.png" alt="CMO Logo" width="500"> </p> <p align="center"> 🤗 <a href="https://huggingface.co/Bruece/FLUX.1-dev-CMO-HPSv2"><b>Hugging Face</b></a> &nbsp;|&nbsp; 🌐 <a href="https://thedarkknight-21th.github.io/CMO/"><b>Project Page</b></a> &nbsp;|&nbsp; 🐙 <a href="https://github.com/TheDarkKnight-21th/CMO"><b>GitHub</b></a> &nbsp;|&nbsp; 📄 <a href="https://arxiv.org/abs/2603.18528"><b>arXiv</b></a> </p>

🌟 Official LoRA Adapter for Correlation-Weighted Multi-Reward Optimization for Compositional Generation, accepted to ECCV 2026

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.

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

<details> <summary>ConceptMix (<a href="https://arxiv.org/abs/2408.14339">Link</a>)</summary> <br> <img src="./conceptmix_results.png" alt="ConceptMix Results"> </details> <details> <summary>GenEval 2 (<a href="https://arxiv.org/abs/2512.16853">Link</a>)</summary> <br> <img src="./GenEval2_results.png" alt="GenEval 2 Results"> </details> <details> <summary>T2I-CompBench (<a href="https://arxiv.org/pdf/2307.06350v2">Link</a>)</summary> <br> <img src="./T2I-CompBench_results.png" alt="T2I-CompBench Results"> </details>

🛠️ 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:

@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}
}