> to recreate said object given outside changes... Unless we're talking about something other than diffusion models, the new image is generated from initial ra

> to recreate said object given outside changes... Unless we're talking about something other than diffusion models, the new image is generated from initial random noise. It also doesn't "recreate" parts of it's training data, unless the model was overfitted to a particular work, or the prompt is deliberately designed to do so, or the user is running it in img2img mode against a preexisting work.