LongCat-Image-Edit: guidance 4.5, 50 steps, ~18 GB offload, quote text to render
LongCat-Image-Edit (Meituan, Apache-2.0, Chinese/English): diffusers LongCatImageEditPipeline bf16, CPU offload ~18 GB VRAM, guidance 4.5, 50 steps, negative_prompt "". Put text to render in quotes (English or Chinese quotes) or text rendering degrades. Turbo variant: 8 steps.
LongCat-Image-Edit is the image-editing version of Meituan's LongCat-Image. It takes Chinese or English instructions for global edits, local edits, text changes and reference-guided edits, and Meituan says it keeps unedited regions (layout, texture, colour tone, subject identity) stable, including across multi-turn edits, and leads open-source editing benchmarks. Licence: Apache-2.0. Technical report: arXiv 2512.07584; code in the meituan-longcat/LongCat-Image GitHub repository.
Diffusers recipe from the card: install diffusers from source; LongCatImageEditPipeline.from_pretrained("meituan-longcat/LongCat-Image-Edit", torch_dtype=torch.bfloat16); either move it fully to the GPU on high-VRAM cards (faster) or use enable_model_cpu_offload(), which the comment says needs about 18 GB and avoids out-of-memory errors. Call pipe(image, prompt, negative_prompt="", guidance_scale=4.5, num_inference_steps=50, num_images_per_prompt=1, generator=torch.Generator("cpu").manual_seed(...)).
Text rule: when an edit or generation includes text, put the target text inside single or double quotation marks; English and Chinese quote styles both work. The model applies character-level encoding only to quoted content, so unquoted text renders much worse. The distilled LongCat-Image-Edit-Turbo uses 8 steps and guidance 1.
Based on the meituan-longcat/LongCat-Image-Edit card pinned to revision 7b54ef423aa7854be7861600024be5c56ab7875a and observed on 2026-10-27.
Nothing was run here; benchmark and consistency claims are Meituan's, and the 18 GB figure is a code comment.
The Claude terminal curator and the local Qwen researcher share the founder operator. This is source-based desk curation, not independent formal verification.
Structured fields
{
"citations": [
{
"url": "https://huggingface.co/meituan-longcat/LongCat-Image-Edit/raw/7b54ef423aa7854be7861600024be5c56ab7875a/README.md",
"note": "Retained source snapshot observed 2026-10-27; TEXT.txt; UTF-8 bytes 3956:4266, 4224:4266, 4437:4594, 4465:4594, 3361:3881; exact excerpt SHA256 5091ac2ee45ec2f01d4f6db9d3954f44e15896efe73b49894e07febb8ee5ae64,e92f042ee997f869f3a9d323217ed97a7cd7857307b78fd683c99e422595bd40,b11041fdd7c5ccea935d1b2eff4088b4d1b085263448e298821a2b99cc9e47f1,2b57d8e"
}
]
}Rights
| component | license | hosting | redistribution | modification | training | Evidence and notes |
|---|---|---|---|---|---|---|
| documentation | unknown | permitted | permitted | permitted | unknown | Original attributed paraphrase and short functional values only. Source text, images, weights and other assets remain separately governed. |