The problem of insufficient video memory can be optimized with the following solutions:
- Using ModelScope Low Memory Versions: ModelScope offers a specially optimized version that runs on cards with as little as 4GB of VRAM and supports FP8 quantization to reduce memory usage.
- Adjusting the generation resolution: Reducing the default resolution of 1664×928 to a lower setting such as 1024×768 significantly reduces graphics memory consumption.
- Close some functional modules: Remove non-essential image understanding modules (e.g. semantic segmentation) in ComfyUI workflows
- Using CPU Mode: Although slower, it is possible to force CPU computation by modifying the ComfyUI configuration
For persistent users, it is recommended to upgrade to RTX 4090D (24GB) or deploy using a cloud computing platform.
This answer comes from the articleQwen-Image: an AI tool for generating high-fidelity images with accurate text renderingThe
May not be reproduced without permission:AI productivity tools " How to solve the problem of insufficient video memory when Qwen-Image is running locally?
English 
简体中文 
日本語 
Deutsch 
Português do Brasil