RefineEdit introduces a training‑free prompt‑to‑prompt image editing framework that leverages a Generative Refinement Network to iteratively refine edits while preserving unrelated source content. It overcomes the spatial inaccuracies of diffusion‑based editors and the fixed decoding order constraints of causal autoregressive methods. The approach enables precise, controllable edits without additional training.