Denoised and original latents are blended regardless of model used. Is this intentional?

[manurare]

Hi,

apply_blend() function gets called always that a mask is present. I think this is OK for models that are not conditioned by concatenation with extra signals or where the unet number of channels is 4 (case of stable diffusion).

However, for models that condition the latents with the masked image latents + mask: why do we need to blend denoised latents with original latents? Isn't this already happening somehow thanks to the conditioning?