Qwen-Image-2.1 releases a 7B-parameter Single-Stream DiT that unifies text-to-image, multi-reference editing up to 10 images, native RGBA transparency, and local region editing, with 33.1GB BF16 weights under a non-commercial research license and a self-reported 60.28 benchmark score.
Grok Image 2.0, released on August 7, adds precise local editing, multi‑reference composition, Smart Resize and templating to its Quality Mode, and while it shows strong visual layout capabilities, third‑party tests reveal weaknesses in Chinese text accuracy, style stability and fine‑grained edits, prompting a cautious but enthusiastic adoption for early‑stage image production.
This guide explains how to unlock GPT‑Image‑2’s four advanced capabilities—multi‑round iteration, natural‑language local editing, multi‑image generation, and reference‑image mode—by showing concrete prompts, code snippets, best‑practice formulas, performance data, and common pitfalls to avoid.
The paper introduces ABPN, an Adaptive Blend Pyramid Network that achieves precise, high‑quality skin retouching and garment wrinkle removal on 4K‑8K photos in real time by combining a context‑aware local retouching layer with a novel adaptive blend pyramid layer, addressing challenges of artifact‑free detail preservation and efficient high‑resolution processing.
