README.md

---
base_model:
- stable-diffusion-v1-5/stable-diffusion-v1-5
license: apache-2.0
pipeline_tag: text-to-image
library_name: diffusers
---

<meta name="google-site-verification" content="-XQC-POJtlDPD3i2KSOxbFkSBde_Uq9obAIh_4mxTkM" />




<div align="center">
  
<h2>DynamicID: Zero-Shot Multi-ID Image Personalization with Flexible Facial Editability</h2>
<h3>[ICCV 2025]</h3>

[Xirui Hu](https://openreview.net/profile?id=~Xirui_Hu1),
[Jiahao Wang](https://openreview.net/profile?id=~Jiahao_Wang14),
[Hao Chen](https://openreview.net/profile?id=~Hao_chen100),
[Weizhan Zhang](https://openreview.net/profile?id=~Weizhan_Zhang1),
[Benqi Wang](https://openreview.net/profile?id=~Benqi_Wang2),
[Yikun Li](https://openreview.net/profile?id=~Yikun_Li1),
[Haishun Nan](https://openreview.net/profile?id=~Haishun_Nan1),

[![arXiv](https://img.shields.io/badge/arXiv-2503.06505-b31b1b.svg)](https://arxiv.org/abs/2503.06505)
[![GitHub](https://img.shields.io/badge/GitHub-MTVCrafter-blue?logo=github)](https://github.com/ByteCat-bot/DynamicID)
</div>

---
This is the official implementation of DynamicID, a framework that generates visually harmonious image featuring **multiple individuals**. Each person in the image can be specified through user-provided reference images, and most notably, our method enables **independent control of each individual's facial expression** via text prompts. Hope you have fun with this demo!

---

## 🔍 Abstract

Recent advancements in text-to-image generation have spurred interest in personalized human image generation. Although existing methods achieve high-fidelity identity preservation, they often struggle with **limited multi-ID usability** and **inadequate facial editability**. 

We present DynamicID, a tuning-free framework that inherently facilitates both single-ID and multi-ID personalized generation with high fidelity and flexible facial editability. Our key innovations include: 

- Semantic-Activated Attention (SAA), which employs query-level activation gating to minimize disruption to the original model when injecting ID features and achieve multi-ID personalization without requiring multi-ID samples during training. 

- Identity-Motion Reconfigurator (IMR), which applies feature-space manipulation to effectively disentangle and reconfigure facial motion and identity features, supporting flexible facial editing.

- A task-decoupled training paradigm that reduces data dependency

- A curated VariFace-10k facial dataset, comprising 10k unique individuals, each represented by 35 distinct facial images. 

Experimental results demonstrate that DynamicID outperforms state-of-the-art methods in identity fidelity, facial editability, and multi-ID personalization capability.

## 💡 Method

<div align="center">
    <img src="assets/pipeline.jpg", width="1000">
</div>

The proposed framework is architected around two core components: SAA and IMR. (a) In the anchoring stage, we jointly optimize the SAA and a face encoder to establish robust single-ID and multi-ID personalized generation capabilities. (b) Subsequently in the reconfiguration stage, we freeze these optimized components and leverage them to train the IMR for flexible and fine-grained facial editing.

## 🚀 Checkpoint

1. Download the pretrained Stable Diffusion v1.5 checkpoint from [Stable Diffusion v1.5 on Hugging Face](https://huggingface.co/stable-diffusion-v1-5/stable-diffusion-v1-5).

2. Download our SAA-related and IMR-related checkpoints from  [DynamicID Checkpoints on Hugging Face](https://huggingface.co/meteorite2023/DynamicID).

## ⚡ Sample Usage (Diffusers)

The official inference code is available in the [GitHub repository](https://github.com/ByteCat-bot/DynamicID), which provides detailed instructions for running the model. A typical usage with the `diffusers` library would involve loading the base Stable Diffusion pipeline and then integrating the DynamicID specific weights.

```python
import torch
from diffusers import StableDiffusionPipeline

# Load the base Stable Diffusion pipeline
# Ensure you have downloaded the base model locally or from Hugging Face Hub
pipeline = StableDiffusionPipeline.from_pretrained(
    "runwayml/stable-diffusion-v1-5", torch_dtype=torch.float16
).to("cuda")

# Load DynamicID specific weights (e.g., LoRAs or custom UNet modifications)
# The precise method for loading these weights will be detailed in the official repository.
# For conceptual understanding, it might involve:
# pipeline.load_lora_weights("path/to/DynamicID/weights")
# Or integrating custom UNet/attention layers as per the DynamicID implementation.

# Refer to the official GitHub repository for the exact loading and inference pipeline.
# You would then pass your text prompt and identity reference images to the pipeline.
# Example (conceptual):
# prompt = "a photo of [person1] with a big smile and [person2] looking thoughtful"
# generated_image = pipeline(
#     prompt=prompt,
#     identity_references=[id_image_1, id_image_2], # Placeholder for identity images
#     # Add other parameters as specified in the DynamicID code
# ).images[0]
```

## 🌈 Gallery

<div align="center">
    <img src="assets/teaser.jpg", width="900">
    <br><br><br>
    <img src="assets/single.jpg", width="900">
    <br><br><br>
    <img src="assets/multi.jpg", width="900">
</div>

## 📌 ToDo List

- [x] Release technical report
- [x] Release **training and inference code**
- [x] Release **Dynamic-sd** (based on *stable diffusion v1.5*)  
- [ ] Release **Dynamic-flux** (based on *Flux-dev*)
- [ ] Release a Hugging Face Demo Space

## 📖 Citation
If you are inspired by our work, please cite our paper.
```bibtex
@inproceedings{dynamicid,
      title={DynamicID: Zero-Shot Multi-ID Image Personalization with Flexible Facial Editability},
      author={Xirui Hu,
              Jiahao Wang,
              Hao Chen,
              Weizhan Zhang,
              Benqi Wang,
              Yikun Li,
              Haishun Nan
              },
      booktitle={International Conference on Computer Vision},
      year={2025}
    }
    
```