dnoise bottom of kSampler also helps make image follow prompt instructions more closely

When doing image to image you NEED vae output from model, to vae of VAE Encode of reference images, then the Latent output of VAE Encode to KSAmpler latent_image

Lower denoise, closer to the original image