Windows 10 Privacy Settings (part of the guide but in different location) The Page File, SuperFetch, and ReadyBoostĪdd a Windows Mobility Center shortcut to your desktopĪccessing Windows Recovery Settings Offilne Using The Windows Store With A Local Account Optimize – get the most out of, make optimal – modify to achieve maximum efficiency in storage capacity or time or cost in order to make something more effective or correct. Release 1809 does not have enough significant changes to warrant a rewrite. Guide will not be updated for the October Update. Updated for Windows 10 April 2018 Update. Updated version here: Ī guide to tweak and optimize Windows 10.
I've run it on RunPod and it should work on HuggingFace as well, but you may want to convert the models ahead of time and copy them up/from S3.************ This version is now expired. There are containers available for CPU, CUDA, and ROCm - I couldn't find the right packages for a DirectML container. There are a few downsides, FP32 memory usage is a little bit higher and you have to convert the SD models to ONNX ahead of time, but you can load the LoRAs and Textual Inversions directly from the safetensors, no conversion needed there. Once a model has been blended and loaded, it stays in VRAM and can be pretty fast. It also supports hardware acceleration for SD upscaling, which should land upstream in the next release of diffusers: I found a way to blend the LoRA and Textual Inversion models in-memory so you can include them in the prompt, with different weights, and tweak them as you go. This is using ONNX for hardware acceleration, so it works on all combinations of AMD and Nvidia, Windows and Linux, FP16 and FP32 - even FP16 on AMD on Windows. It has a tab for each mode, persistent and private image history using localStorage, and some basic internationalization (although you can override the language detection using ?lng=en in the URL).
The GUI is included with the bundle or hosted on Github Pages and runs in most modern browsers, including on mobile: You can download and convert models from the HuggingFace hub and Civitai: The API needs Python 3, a few libraries, and about 50GB of disk space to get started (although it will share models from the HuggingFace cache, if you have one). I've been working on another UI for Stable Diffusion on AMD and Windows, as well as Nvidia and/or Linux, and recently added support for blending LoRA and Textual Inversions using the prompt: Įdit : based on feedback that my previous setup instructions were difficult and error-prone, I've added a few more ways to get started, including an all-in-one bundle for Windows:
0 kommentar(er)
