There has been significant progress in personalized image synthesis with methods such as Textual Inversion, DreamBooth, and LoRA. Yet, their real-world applicability is hindered by high storage demands, lengthy fine-tuning processes, and the need for multiple reference images. Conversely, existing ID embedding-based methods, while requiring only a single forward inference, face challenges: they either necessitate extensive fine-tuning across numerous model parameters, lack compatibility with community pre-trained models, or fail to maintain high face fidelity. Addressing these limitations, we introduce InstantID, a powerful diffusion model-based solution. Our plug-and-play module adeptly handles image personalization in various styles using just a single facial image, while ensuring high fidelity. To achieve this, we design a novel IdentityNet by imposing strong semantic and weak spatial conditions, integrating facial and landmark images with textual prompts to steer the image generation. InstantID demonstrates exceptional performance and efficiency, proving highly beneficial in real-world applications where identity preservation is paramount. Moreover, our work seamlessly integrates with popular pre-trained text-to-image diffusion models like SD1.5 and SDXL, serving as an adaptable plugin. Our codes and pre-trained checkpoints will be available at this URL.
Given only one reference ID image, InstantID aims to generate customized images with various poses or styles from a single reference ID image while ensuring high fidelity. Following figure provides an overview of our method. It incorporates three crucial components: (1) An ID embedding that captures robust semantic face information; (2) A lightweight adapted module with decoupled cross-attention, facilitating the use of an image as a visual prompt; (3) An IdentityNet that encodes the detailed features from the reference facial image with additional spatial control.
We are different from previous works in the following aspects: (1) We do not train UNet, so we can preserve the generation ability of the original text-to-image model and be compatible with existing pre-trained models and ControlNets in the community; (2) We don't require test-time tuning, so for a specific character, there is no need to collect multiple images for fine-tuning, only a single image needs to be inferred once; (3) We achieve better face fidelity, and retain the editability of text.
InstantID supports both Stylizated and Realistic styles. Scroll for more examples.
Demonstration of the robustness, editability, and compatibility of InstantID. Column 1 shows the result of Image Only results where the prompt is set to empty during inference. Columns 2-4 show the editability through text prompt. Columns 5-9 show the compatibility with existing ControlNets (canny & depth).
Effect of the number of reference images. For multiple reference images, we take the average mean of ID embeddings as image prompt. InstantID is able to achieve good results even with only one single reference image.
Comparison with existing tuning-free state-of-the-art techniques. Specifically, we compare with IP-Adapter (IPA), IP-Adapter-FaceID, and recent PhotoMaker. Among them, PhotoMaker needs to train the LoRA parameters of UNet. It can be seen that both PhotoMaker and IP-Adapter-FaceID achieves good fidelity, but there is obvious degradation of text control capabilities. In contrast, InstantID achieves better fidelity and retain good text editability (faces and styles blend better).
Comparison of InstantID with pre-trained character LoRAs. We can achieve competitive results as LoRAs without any training.
Comparison of InstantID with InsightFace Swapper (also known as ROOP or Refactor). However, in non-realistic style, our work is more flexible on the integration of face and background.
Interpolation between two different characters.
Our work also flexiblely supports add identity attribute into a non-human character.
@article{wang2024instantid,
title={InstantID: Zero-shot Identity-Preserving Generation in Seconds},
author={Wang, Qixun and Bai, Xu and Wang, Haofan and Qin, Zekui and Chen, Anthony},
journal={arXiv preprint arXiv:2401.07519},
year={2024}
}