search

SHI-Labs / Smooth-Diffusion

Smooth Diffusion: Crafting Smooth Latent Spaces in Diffusion Models arXiv 2023 / CVPR 2024
https://shi-labs.github.io/Smooth-Diffusion/
MIT License
313 stars 8 forks source link

Why is text embedding linear interpolated? #18

Open wtliao opened 6 days ago

wtliao commented 6 days ago

Many thanks for sharing your impressive work and code. I am reading your code line by line for better understanding. One point I cannot figure out is that, why is text embedding also linearly interpolated (emb_list = auto_lerp(tlist, emb0, emb1) )? In my opinion, this kind of interpolated will change the condition (i.e. prompt), and all initial noisy input should share the same prompt condition. Please help me understand it. Many thanks again!

JiayiGuo821 commented 6 days ago

In the paper cases, two input images for interpolation share the same prompt, i.e., emb0=emb1. Therefore, the linear interpolation makes no difference for these cases since (1-a)*emb+a*emb=emb. We provide this prompt linear interpolation for potential user testing for interpolating images with different prompts.