Meituan open sources LongCat-Image-Edit-Turbo, a distilled image editing model that hits open source SOTA in only 8 inference steps

Meituan's LongCat team has open-sourced LongCat-Image-Edit-Turbo, a distilled version of their instruction-based image editing model. The key breakthrough is its efficiency: it matches the quality of its larger predecessor in just 8 NFEs (Number of Function Evaluations), representing a roughly 10x speedup. This performance is achieved with a compact 6B-parameter diffusion core, requiring only about 18GB of VRAM with CPU offloading. The model is fully integrated into the HuggingFace Diffusers library and is Apache 2.0 licensed, making it immediately accessible for developers and researchers.

On standard benchmarks, LongCat-Image-Edit-Turbo sets a new open-source SOTA. It scores 4.50 on ImgEdit-Bench and 7.60/7.64 on GEdit-Bench for Chinese and English instructions, respectively, outperforming competitors like FLUX.1 Kontext and Qwen-Image-Edit. Its editing capabilities are extensive, covering global and local edits, object replacement, style transfer, text manipulation, and outpainting. A standout feature is its exceptional consistency preservation, which maintains the layout, texture, and identity of non-edited regions—a critical factor for professional multi-turn editing workflows.

The model's efficiency stems from rigorous data curation and distillation techniques, continuing a trend where smart training methods outperform brute-force parameter scaling. It natively supports both Chinese and English instructions and includes a clever character-level encoding trick for accurate text rendering. With ComfyUI support and training code also released, Meituan's release demonstrates how well-trained, mid-sized models can deliver top-tier performance at a fraction of the computational cost, pushing the boundaries of what's possible in open-source AI.