Closed miangoleh closed 1 month ago
Hi Miangoleh!Thank you for your interests! StableDelight is just an open-source project and we don't have plan to write a technical report yet. It builds upon the first-stage YOSO from StableNormal, with different training data and additional loss function. The training data include Hypersim, Lumos, and various Specular Highlight Removal datasets from TSHRNet. Also, we've integrated a multi-scale SSIM loss into our diffusion training process, to improve sharpness in one-step diffusion prediction.
We're currently in the optimization phase and would greatly value any feedback!
Hi!
I was wondering if there is a technical report or an accompanying paper that describes the underlying algorithm in more detail.