Views
No views yet
Qwen-Image-Edit-AIO-FP8 is an FP8-compressed edition of the Qwen-Image-Edit series developed by the Qwen team at Alibaba, designed to deliver high-precision multimodal diffusion-based image editing with significantly reduced memory consumption and accelerated inference performance. Built upon the advanced capabilities introduced in Qwen-Image-Edit-2509, including native multi-image conditioning that enables seamless fusion of multiple references such as identity, product, and environment within a single coherent generation pipeline, and further enhanced by the large-scale MMDiT-based architecture of Qwen-Image-Edit-2511 with its 20B-parameter backbone for superior identity consistency and minimal drift during iterative refinement, this FP8 release preserves structural control, text-in-image editing accuracy, and ControlNet compatibility while dramatically improving deployment efficiency on modern hardware. Optimized for industrial design workflows, high-fidelity multi-person composition, material replacement, geometric reasoning, and annotation-aware generation, Qwen-Image-Edit-AIO-FP8 maintains professional-grade lighting and viewpoint stability, supports integrated community LoRA adaptations, and enables scalable production-ready image editing pipelines with lower VRAM requirements and minimal quality degradation compared to full-precision checkpoints.
1transformers # - transformers@v4.57.6
2torch # - torch@v2.9.1+cu128
3diffusers # - diffusers@v0.37.0.dev0
4accelerate # - accelerate@v1.12.01import torch
2from diffusers import QwenImageEditPipeline # or your compatible pipeline
3
4model_path = "./Qwen-Image-Edit-AIO-FP8"
5
6pipe = QwenImageEditPipeline.from_pretrained(
7 model_path,
8 torch_dtype=torch.bfloat16
9)
10
11pipe.to("cuda")
12
13# Example usage
14# outputs = pipe(image=input_image, prompt="Your edit prompt")[!IMPORTANT] This repository follows the same release notes, terms and conditions, and license as the original model page, Qwen-Image-Edit.
[!IMPORTANT] This repository also contains experimental LLM compressed (compressed-tensors) editions of the Qwen image edit models. Feel free to ignore those editions if you are struggling with them or unable to mount them.