We introduce two learnable tokens in the prompt to generate complementary cross-attention maps. These tokens are designed to capture both the object and the background concept, which includes all regions outside the objects. We design a two-stage pipeline. Throughout the two stages, we set the prompt as "an aerial view image with [V1] [category] in [V2] [S]" for source domain data and "an aerial view image with [V1] [category] in
[V3] [T]" for target domain data. [V1] represents the learnable token for the object concept, while [V2] and [V3] correspond to the learnable token for the source domain and target domain background concept, respectively.
In the first stage, we fine-tune both the U-Net and learnable tokens [V1], [V2], and [V3] to capture the new concepts.
Model Usage
This folder contains the fine-tuned Stable Diffusion and the learnable embeddings in the first stage.
