![]() There are several cross attention optimization methods such as -xformers or -opt-sdp-attention, these can drastically increase performance see Optimizations for more details, experiment with different options as different hardware are suited for different optimizations. ![]() Some combinations of model and VAE are prone to produce NansException: A tensor with all NaNs was produced in VAE resulting in a black image, using the option -no-half-vae may help to mitigate this issue. If your generated results in resulting in a black or green image, try adding -precision full and -no-half. The Tiled VAE extension can help to reduce the VRAM requirement. The amount of VRAM required largely depends the desired image resolution, for more details see Troubleshooting. If you do not have enough VRAM, web UI may refuse to launch or fail to generate images due to an out-of-memory error.
0 Comments
Leave a Reply. |
Details
AuthorWrite something about yourself. No need to be fancy, just an overview. ArchivesCategories |