We combined OneDiff DeepCache and LoRA, achieving more than a 2.7x acceleration for shapes of 1024*1024, workflow reference:
For the Aether Cloud style of LoRA, it can be downloaded from https://civitai.com/models/141029/aether-cloud-lora-for-sdxl.
Updated on Jan 11, 2024. Device: A100 GPU
| LoRA | LoRA + DeepCache | Percentage improvement | |
|---|---|---|---|
| 512 * 1024 | 16.30 it/s | 39.57 it/s | 142.45% |
| 1024 * 1024 | 9.10 it/s | 25.19 it/s | 176.04% |
However, due to the characteristics of DeepCache, combining it with LoRA introduces some differences in the output:
Note: If your system or business requires more personalized support, please check: https://github.com/siliconflow/onediff?tab=readme-ov-file#onediff-enterprise-edition