In our model's image-conditioned branch, we integrated three variants of the CLIP model with differing backbones to enhance the vision encoder's performance, producing a robust final output.
For the training phase, we relied on the FFHQ face image dataset alongside a subset of the LAION dataset to ensure the model's ability to generate diverse images, striking a balance between human and non-human generation.
Incorporating randomness during training by omitting conditional embeddings allows our model to adaptively handle variability in the generated outputs, optimizing for various styles and identities.
Utilizing the EulerA sampler and implementing classifier-free guidance with strategically omitted embeddings, we tailored the training process to enhance the quality of image generation.
Collection
[
|
...
]