hummingbird

Hummingbird: High Fidelity Image Generation via Multimodal Context Alignment [ICLR 2025]

This repository contains the LoRA weights for the Hummingbird model, presented in Hummingbird: High Fidelity Image Generation via Multimodal Context Alignment. The Hummingbird model generates high-quality, diverse images from a multimodal context, preserving scene attributes and object interactions from both a reference image and text guidance.

Project page | Paper

Official implementation of paper: Hummingbird: High Fidelity Image Generation via Multimodal Context Alignment

Prerequisites

Installation

1. Clone this repository and navigate to hummingbird-1 folder

git clone https://github.com/roar-ai/hummingbird-1
cd hummingbird-1

1. Create conda virtual environment with Python 3.9, PyTorch 2.0+ is recommended:

conda create -n hummingbird python=3.9
conda activate hummingbird
pip install torch==2.4.1 torchvision==0.19.1 torchaudio==2.4.1 --index-url https://download.pytorch.org/whl/cu124
pip install -r requirements.txt

1. Install additional packages for faster training and inference

pip install flash-attn --no-build-isolation

Download necessary models

1. Clone our Hummingbird LoRA weight of UNet denoiser

git clone https://huggingface.co/lmquan/hummingbird

1. Refer to stabilityai/stable-diffusion-xl-base-1.0 to download SDXL pre-trained model and place it in the hummingbird weight directory as ./hummingbird/stable-diffusion-xl-base-1.0.

2. Download laion/CLIP-ViT-bigG-14-laion2B-39B-b160k for feature extractor and image encoder in Hummmingbird framework

cp -r CLIP-ViT-bigG-14-laion2B-39B-b160k ./hummingbird/stable-diffusion-xl-base-1.0/image_encoder

mv CLIP-ViT-bigG-14-laion2B-39B-b160k ./hummingbird/stable-diffusion-xl-base-1.0/feature_extractor

1. Replace the file model_index.json of pre-trained stable-diffusion-xl-base-1.0 with our customized version for Hummingbird framework

cp -r ./hummingbird/model_index.json ./hummingbird/stable-diffusion-xl-base-1.0/

1. Download HPSv2 weights and put it here: hpsv2/HPS_v2_compressed.pt.
2. Download PickScore model weights and put it here: pickscore/pickmodel/model.safetensors.

Double check if everything is all set

|-- hummingbird-1/
    |-- hpsv2
        |-- HPS_v2_compressed.pt
    |-- pickscore
        |-- pickmodel
            |-- config.json
            |-- model.safetensors
    |-- hummingbird
        |-- model_index.json
        |-- lora_unet_65000
            |-- adapter_config.json
            |-- adapter_model.safetensors
        |-- stable-diffusion-xl-base-1.0
            |-- model_index.json (replaced by our customized version, see step 4 above)
            |-- feature_extractor (cloned from CLIP-ViT-bigG-14-laion2B-39B-b160k)
            |-- image_encoder (cloned from CLIP-ViT-bigG-14-laion2B-39B-b160k)
            |-- text_encoder
            |-- text_encoder_2
            |-- tokenizer
            |-- tokenizer_2
            |-- unet
            |-- vae
            |-- ...
    |-- ...

Quick Start

Given a reference image, Hummingbird can generate diverse variants of it and preserve specific properties/attributes, for example:

python3 inference.py --reference_image ./examples/image-2.jpg --attribute "color of skateboard wheels" --output_path output.jpg

Training

You can train Hummingbird with the following script:

sh run_hummingbird.sh

Synthetic Data Generation

You can generate synthetic data with Hummingbird framework, for e.g. with MME Perception dataset:

python3 image_generation.py --generator hummingbird --dataset mme --save_image_gen ./synthetic_mme

Testing

Evaluate the fidelity of generated images w.r.t reference image using Test-Time Augmentation on MLLMs (LLaVA/InternVL2):

python3 test_hummingbird_mme.py --dataset mme --model llava --synthetic_dir ./synthetic_mme

Acknowledgement

We base on the implementation of TextCraftor. We thank BLIP-2 QFormer, HPSv2, PickScore, Aesthetic for the reward models and MLLMs LLaVA, InternVL2 functioning as context descriptors in our framework.

Citation

If you find this work helpful, please cite our paper:

@inproceedings{le2025hummingbird,
    title={Hummingbird: High Fidelity Image Generation via Multimodal Context Alignment},
    author={Minh-Quan Le and Gaurav Mittal and Tianjian Meng and A S M Iftekhar and Vishwas Suryanarayanan and Barun Patra and Dimitris Samaras and Mei Chen},
    booktitle={The Thirteenth International Conference on Learning Representations},
    year={2025},
    url={https://openreview.net/forum?id=6kPBThI6ZJ}
}