Towards a Simultaneous and Granular Identity-Expression Control in Personalized Face Generation

Renshuai Liu1, 2, Bowen Ma2, Wei Zhang2, Zhipeng Hu2, Changjie Fan2, Tangjie Lv2, Yu Ding2*, Xuan Cheng1*
1School of Informatics, Xiamen University, 2Virtual Human Group, Netease Fuxi AI Lab

DiffSFSR takes three inputs: a prompt describing the background, a selfie photo uploaded by the user, and a text related to the fine-grained expression labels. The generated faces well match the inputted triples and exhibit fine-grained expression synthesis.

Abstract

In human-centric content generation, the pre-trained text-to-image models struggle to produce user-wanted portrait images, which retain the identity of individuals while exhibiting diverse expressions.

This paper introduces our efforts towards personalized face generation. To this end, we propose a novel multi-modal face generation framework, capable of simultaneous identity-expression control and more fine-grained expression synthesis. Our expression control is so sophisticated that it can be specialized by the fine-grained emotional vocabulary. We devise a novel diffusion model that can undertake the task of simultaneously face swapping and reenactment. Due to the entanglement of identity and expression, it's nontrivial to separately and precisely control them in one framework, thus has not been explored yet. To overcome this, we propose several innovative designs in the conditional diffusion model, including balancing identity and expression encoder, improved midpoint sampling, and explicitly background conditioning.

Extensive experiments have demonstrated the controllability and scalability of the proposed framework, in comparison with state-of-the-art text-to-image, face swapping, and face reenactment methods.

Video

Expression Travel

DiffSFSR also supports expression travel by interpolating between embeddings to further explore fine-grained expression.

Interpolate start reference image.

Start Frame

Loading...
Interpolation end reference image.

End Frame


Improved Midpoint Sampling

Qualitative

Here are some cases of sampling results. As shown above, Our results are not only more faithful to ground truth, but also more realistic and clear in the regions of eyes, mouths and even the reflection of sunglasses.

Quantitative

We employ MSE to measure the error between sampling results and ground truth. All sampling methods can decrease the reconstruction errors along with the training steps increasing. Our sampling method can achieve lower MSE than others in all periods.

Fine-grained Expression Controlling Results

PDF results

Google Drive
Baidu Netdisk

JPG results

BibTeX


      @article{liu2024simultaneous,
        title={Towards a Simultaneous and Granular Identity-Expression Control in Personalized Face Generation},
        author={Renshuai Liu and Bowen Ma and Wei Zhang and Zhipeng Hu and Changjie Fan and Tangjie Lv and Yu Ding and Xuan Cheng},
        journal={arXiv preprint arXiv:2401.01207},
        year={2024}
      }