there exists challenges in infusing the prior distribution information into the diffusion steps: if directly modify the sample hk at step k (i.e., hk ∼ Hk) without careful consideration, it can potentially cause the sample hk to deviate from the distribution Hk, which disrupts the diffusion process(the mesh distribution Hk at each k-th step of the process has its own characteristics)
Keyideas
infuse prior distribution information(U: pose heatmap) to the mesh distribution diffusion process without disrupting it--->DAT, which aligns the diffusion steps towards the prior distribution by taking the k remaining diffusion steps into account( DDIM ), carefully updating hk such that the eventual prediction after k diffusion steps is pulled closer to the diffusion target.
DAT
Distribution Alignment Gradient+Activation Strategy
-Full architecture
Code available: NO until 4/20