To address the pain point of insufficient consumer GPU graphics memory, Wan 2.2 offers three optimization options:
Option 1: Selection of TI2V-5B model
- This model is specifically optimized for 24GB video memory GPUs (e.g. RTX 4090)
- Run with parameters: -offload_model True -convert_model_dtype -t5_cpu
- Reduced graphics memory requirements through model offloading and precision conversion
Option 2: Enable hybrid expert architecture
- MoE architecture naturally reduces compute load
- Separate high/low noise processing allows dynamic allocation of resources
- Especially suitable for resolutions below 1080P
Option 3: Use the ComfyUI graphical interface
- Visualizing Operations to Avoid Command Line Memory Leaks
- Built-in automatic video memory management module
- Supports real-time monitoring of video memory utilization
This answer comes from the articleWan2.2: Open Source Video Generation Model with Efficient Text and Image to Video SupportThe
May not be reproduced without permission:AI productivity tools " How to solve the problem that consumer GPUs don't have enough video memory to run Wan 2.2?
English 
简体中文 
日本語 
Deutsch 
Português do Brasil