Graphics Memory Optimization Solution for High-Resolution Fixes
The following methods can be used to reduce the GPU load when processing images above 4K:
- Partition repair technology: Split the big picture into multiple 512×512 blocks, process them by Batch, and finally stitch them together with image editing software.
- parameterization: Decreasing the value of LanPaint_cfg_Big (the default value recommended by the model) reduces the amount of computation during detail generation
- Resolution Compression: first downsampling of the original image processing (it is recommended that no less than 1/4 size of the original image), after the repair is completed and then use tools such as ESRGAN super-scoring
- VRAM Management: Enable the -medvram parameter in ComfyUI settings, or use the -lowvram mode to force chunking
advanced skill: For devices with weak graphics card performance, you can use Latent Zoom technology in conjunction to complete the repair in the low-dimensional space before up-sampling. Pay attention to monitor the amount of video memory usage (nvidia-smi command), when occupying more than 90% should be immediately interrupted to adjust the parameters.
This answer comes from the articleLanPaint: A ComfyUI Image Restoration Workflow for Any ModelThe
May not be reproduced without permission:AI productivity tools " How to optimize LanPaint's video memory usage for high resolution image restoration?
English 
简体中文 
日本語 
Deutsch 
Português do Brasil