inference stage layerdiffusion: layer_xl_transparent_attn.safetensors

[Xiaobin HU(kevin)]

Thank you for your commendable work. Your remarkable contributions have served as a significant source of inspiration for recent advancements in layer-based image editing. Figure 2 of your paper on Transparent Image Layer Diffusion using Latent Transparency (https://arxiv.org/pdf/2402.17113.pdf) demonstrates that during the training phase, only the encoder and decoder weights are required, while the remaining weights of sd-xl are held fixed. However, I am currently perplexed regarding the inference stage. Is it necessary to employ the alpha-based encoder weight despite the absence of an alpha channel? As far as my understanding goes, for the sole purpose of foreground image generation during inference, a well-trained decoder and other layer_xl_transparent_attn.safetensors are indispensable. Where does layer_xl_transparent_attn.safetensors come from? how it train in paper. I greatly appreciate your reply.