The paper presents a method to adapt pretrained raster-scan autoregressive (AR) image models into any-order generators capable of inpainting, outpainting, and editing without retraining from scratch. The authors identify that fixed-order training creates a target-position identifiability problem, which they resolve using future-aware positional embeddings that enable the model to condition on arbitrary token subsets. An ordering curriculum provides a stable optimization path from the pretrained raster-scan solution to an any-order basin. Applied to LlamaGen-L/XL and RAR-L/XL across two architectures and scales, the recipe adds only 13–33% compute beyond pretraining while preserving compatibility with standard AR inference infrastructure (KV caching, vLLM). On ImageNet, the adapted models achieve inpainting FID of 6.21 versus MaskGIT's 10.44, outperform MaskGIT on outpainting, and remain competitive with MAR-L. This flexibility trades a modest 1.2–1.9 increase in class-conditional FID compared to the original raster-scan models. The approach demonstrates that editing capabilities can be unlocked from existing pretrained AR models rather than requiring dedicated training from scratch.