Show HN: UniWorld V2 – Region-Aware AI Image Editing with RL Accuracy (www.uniworldv2.com)

UniWorld V2 is an instruction-driven image-editing model that couples a diffusion backbone with reinforcement learning to deliver precise, region-aware edits from natural-language prompts. The system uses an RL framework (Edit-R1 / UniWorld-R2) where a multi-modal large language model (MLLM) evaluates edit quality as a reward signal, improving alignment to user intent, structural consistency, and instruction correctness. Practically, users draw a mask (rectangle or polygon), type an instruction (e.g., “replace the bag with a red handbag” or “make the font calligraphy style”), and UniWorld V2 applies edits only within the selected region while preserving global lighting, perspective and composition. The team reports it outperforms models like GPT-Image-1, Nano Banana and Gemini on edit-accuracy benchmarks. Its standout capabilities are stable multi-round editing (edit → re-edit → refine without breaking style), and advanced typography handling: text is treated as a first-class visual element—maintaining font, strokes, spacing and perspective instead of being smeared as texture. Examples include moving/removing objects, changing gestures, extracting elements, and full scene re-composition with coherent lighting. The result is a practical tool for advertising, localization, UI/UX iteration, publishing and e‑commerce workflows where rapid, faithful visual changes across multiple rounds are required.