In a preprint posted to arXiv, researchers argue that instruction-driven image editing at native 4K cannot be solved by a single end-to-end editor. Users write vague instructions, and precise parametric or specialist edits are better handled by combining a core model with an agent that plans, verifies, and retries. Their proposed system, BluePixel, couples a native-4K editing model with such an agent harness, and the paper discloses the complete training recipe.
The training pipeline turns a 1K-pretrained editor into a native-4K one through progressive VAE adaptation, which keeps 4K training as affordable as 1K training. Continual learning absorbs a long tail of vertical editing skills, and reinforcement learning refines the trade-off between aesthetics and fidelity. At inference, a distillation method called source-anchored target projection compresses 50 denoising steps down to 4, completing a 4K edit in 8.7 seconds on a single Kunlunxin P800 XPU while preserving task success and background consistency.
The agent component decomposes user instructions into specialist and general operations, checks every result with the same reward model used during training, and retries within bounded attempts. On VINS-4KEval++, a benchmark of 3,305 native-4K cases, BluePixel ranked second among seven models, trailing GPT Image 2 only in how consistently it preserved non-edited regions. The authors position BluePixel as evidence that the path to practical 4K editing lies in open, system-level designs rather than monolithic model scaling.