Assuming you're using diffusion models, just decode them with VAEs back into pixel space, you know they're compressed into latent space to save rss... or do not use latent imgs at all it's slower but you can access the imgs anytime