UnZipLoRA: Separating Content and Style from a Single Image

[Paper] [Project Page]

(ICCV 2025 Highlight)

This repository contains the official implementation of UnZipLoRA -- a novel technique for decoupling content and style from single input image by learning two distinct LoRAs simultaneously. UnZipLoRA ensures that the resulting LoRAs are compatible, i.e., they can be seamlessly combined using direct addition. UnZipLoRA enables independent manipulation and recontextualization of subject and style, including generating variations of each, applying the extracted style to new subjects, and recombining them to reconstruct the original image or create novel variations. For more details, please refer to our paper UnZipLoRA: Separating Content and Style from a Single Image.

Requirements

Install dependencies:

conda create -n unziplora python=3.11
pip install torch==2.3.0 torchvision==0.18.0 torchaudio==2.3.0 --index-url https://download.pytorch.org/whl/cu118
conda activate unziplora
pip install -r requirements.txt
pip install jupyter notebook

Data

We provide the figures and corresponding prompts used for training and evaluation in the data.

Train

To train the content and style model symotanenously, you need to provide:

• One or more reference images
• A prompt, a content prompt and a style prompt
• Output path to save the trained models

You can directly run the training script.

UnZipLoRA proposes three separation strategies. All hyperparameters in the script match the settings used in the paper, and can be adjusted for ablation and tuning.

Core Hyperparameters

Flag	Description	Default
--pretrained_model_name_or_path	Path to base model (e.g. SDXL)	Required
--instance_data_dir	Path to training images	Required
--output_dir	Path to save outputs	Required
--instance_prompt	Prompt used for training	Required
--content_forward_prompt	Prompt used for training(content)	Required
--style_forward_prompt	Prompt used for training(style)	Required
--rank	LoRA rank	64
--max_train_steps	Number of training steps	600
--content_learning_rate	LR for content LoRA	5e-5
--style_learning_rate	LR for style LoRA	5e-5
--weight_learning_rate	LR for weight fusion	5e-3

---

Logging / Ablation Flags

Flag	Description	Default
--with_period_column_separation	Apply column separation	True
--with_freeze_unet	Apply block separation	True
--with_saved_per_validation	Save model on every validation step	False
--with_image_per_validation	Generate images at every validation step	False
--with_grad_record	Visualize the cone heatmap for selected layers	False

---

Experimental

Flag	Description	Default
--sample_times	Times of re-caliberate LoRA column masks	3
--column_ratio	Ratio of sampled columns for LoRA mask	0.1

Infer

After training, use infer script to generate images with your trained LoRAs. You can:

• Generate content / style recontextualizations with individual saved content / style models and given prompts
• Generate combined recontexualization by merging both content and style LoRAs with learned mask weights
• Generate cross-image recontexualization by combining LoRAs and mask weights from different source figures

We also provide a notebook to play with these functionalities.

You can download two example models from Google Drive and try out these functionalities using them.

Citations

If you find this work helpful, please cite us as:

@InProceedings{Liu_2025_ICCV,
    author    = {Liu, Chang and Shah, Viraj and Cui, Aiyu and Lazebnik, Svetlana},
    title     = {UnZipLoRA: Separating Content and Style from a Single Image},
    booktitle = {Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV)},
    month     = {October},
    year      = {2025},
    pages     = {16776-16785}
}