Table of Links

Abstract and 1 Introduction

2. Related Work

3. Method and 3.1. Hybrid Guidance Strategy

3.2. Handling Multiple Identities

3.3. Training

4. Experiments

4.1. Implementation details.

4.2. Results

5. Conclusion and References

4. Experiments

4.1. Implementation details.

The vision encoder utilized in the image-conditioned branch of our model combines three CLIP model [40] variants with different backbones. These are: CLIP-ViT-L/14, CLIP-RN101, and CLIP-ViT-B/32. The outputs from these individual models are concatenated to produce the final output of our vision encoder. Our approach primarily utilizes the DDPM configuration [20] as described in StableDiffusion [42] for training. Specifically, we incorporated a total of 1,000 denoising steps. For the inference stage, we use the EulerA sampler [2] and set it to operate over 25 timesteps. To align with the training methodology of classifier-free guidance [19], we introduced variability by randomly omitting the conditional embeddings related to both style images and face images. Specifically, the probabilities for dropping these embeddings were set at 0.64 for style images and 0.1 for face images.

The primary dataset used for training was FFHQ [25], which is a face image dataset encompassing 70,000 images. To augment this, we also incorporated a subset of the LAION dataset [46] into our training phase, which aims to ensure the model retains the capability to generate generic, non-human images during the finetuning process. It’s worth noting that when non-human images are sampled for training, the face embedding in the conditional branch is set to zero. During training, we set the learning rate at 1e-6. The model was trained using 8 A100 GPUs, with a batch size of 256, and was trained for 100,000 steps.