LongCat-Image-Edit

Overview

LongCat-Image-Edit takes one existing picture and a written instruction, and returns the picture with that change applied. Meituan LongCat built it from a mid-training checkpoint of its 6B LongCat-Image text-to-image model. The two stay separate because, when the team trained one model for both jobs, the synthetic editing data made text-to-image output less photorealistic.

Anything the instruction does not mention, such as layout, texture, color tone or a person's identity, is meant to stay fixed, even over several rounds of edits. In the samples here, a tabby cat on a dark grey sofa becomes a golden dog with the cushions and lamp untouched, and a yellow pickup turns red while the log cabin and peaks behind it stay put.

Meituan reports 7.64 on GEdit-Bench-EN and 7.67 on its own CEdit-Bench-EN, the best of the open models in its tables. In side-by-side human ratings for comprehensive quality, its win rate was 60.5% against FLUX.1 Kontext [Pro] and 58.7% against Qwen-Image-Edit [2509], with each tie counted as half a win.

What you can build

• Recoloring an object in a product or lifestyle photo, as with the pickup that goes from yellow to red
• Object removal and insertion, such as Meituan's example that removes a laptop and adds a mug
• Portrait retouching: Meituan demonstrates blemish removal, hairstyle changes, relighting and ID photos
• New wording on an existing sign or label
• Background swaps and style transfer on photos you already have

Inputs

Each run takes one source image (JPG, PNG, GIF, WebP or HEIC) and an instruction in English or Chinese, such as "Turn the yellow car into a red car." When the edit involves lettering, put the new words inside single or double quotes; Chinese quotation marks work too.

The advanced panel holds four optional settings. Steps run from 1 to 50 and start at 28. Guidance runs from 1 to 20, defaulting to 4.5, the value in Meituan's own editing example. You can request up to 8 versions at once, and a fixed seed makes a run repeatable. There is no aspect ratio setting.

Outputs

The same scene with the requested change applied. Every sample pair on this page goes in and comes out at 1024x1024, framed identically.

Recommended settings

• Name the object and the result you want, the way the default prompt goes from "the yellow car" to "a red car".
• Several changes can be chained one per run or written as a single instruction. Meituan's report shows a six-step sequence and the combined version ending up close to each other.
• Keep the seed fixed when comparing two wordings of the same instruction.

Limitations

• Only one image goes in per run, and there is no mask, so the instruction itself has to identify what should change.
• In the team's human study, Nano Banana had a 60.8% win rate against it on comprehensive quality and Seedream 4.0 had 56.9%, counting each tie as half a win. Both also came out ahead on consistency, at 53.9% and 56.3%.
• GPT Image 1 [High] leads Meituan's CEdit-Bench-EN table at 8.17.

Safety & compliance

Meituan publishes LongCat-Image-Edit under Apache 2.0. If the photo you upload shows a real person's face, edit it only with that person's consent. The project's usage notes make whoever deploys the model responsible for following the law, including privacy and data protection rules.