base_model:
- stable-diffusion-v1-5/stable-diffusion-v1-5
license: apache-2.0
pipeline_tag: text-to-image
library_name: diffusers
DynamicID: Zero-Shot Multi-ID Image Personalization with Flexible Facial Editability
[ICCV 2025]
Xirui Hu, Jiahao Wang, Hao Chen, Weizhan Zhang, Benqi Wang, Yikun Li, Haishun Nan,
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
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
Download the pretrained Stable Diffusion v1.5 checkpoint from Stable Diffusion v1.5 on Hugging Face.
Download our SAA-related and IMR-related checkpoints from DynamicID Checkpoints on Hugging Face.
⚡ Sample Usage (Diffusers)
The official inference code is available in the GitHub repository, 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.
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
📌 ToDo List
- Release technical report
- Release training and inference code
- 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.
@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}
}