SpecifyGradient is not straghtforward, use a reparameterization trick instead

[YoujiaZhang]

Thanks for the great job for prolificdreamer and DreamFusion! I have a question about the calculation of SDS loss. I found in the paper g = matmul(weight(t) * (-epshat t - eps), grad(x, params)).

threestudio uses this formula d(loss)/d(latents) = latents - target = latents - (latents - grad) = grad when calculating SDS.

grad = weight(t) * (-epshat t - eps) So is this not calculating grad(x, params) ? I seem to be confused by this. :)

Thanks!

[YoujiaZhang]

I solved it :)