Out of curiosity, I attempted to reinstall stable diffusion. I had tried it once before in May 2023. In just a short span of 7 months, there have been many developments and changes, especially regarding the limitations on free resources in Colab. The web UI I originally used was challenging to complete the installation. As a result, I unexpectedly discovered a testing program that significantly enhanced my understanding of Colab and stable diffusion.
Due to the desire to switch models, I found technical documentation and resource management on Hugging Face related to Stable Diffusion. Many developers include test examples in the resource descriptions, providing more learning templates. Although some learning templates may require debugging, it is usually a matter of developers forgetting what tools were installed, and it can be gradually patched during the execution process in Colab. I attempted to install the recently popular SDXL, which functions as img-to-img. However, the GPU provided by Colab for trial accounts only reaches 16GB, while SDXL requires 20GB. Consequently, I opted for the second favorite style, which operates as text-to-img.
Tosh Velaga shared his sample program. The information about "cache_dir" mentioned in it is very useful to me. When I was trying to install the Web UI, I set everything that could be stored to Google Drive to comply with Google's resource limits. When a session is interrupted, I only need to reinstall the VM, saving at least half of the time. I also noticed the need to choose the appropriate runtime type, providing sufficient GPU computing power.
Developers in the Playground mainly use their own trained models, so the code is relatively straightforward. Below is the organized structure of the code for convenient use. It is recommended to obtain the source code from Hugging Face.
Here are test examples I created, using sentences similar to those found in novels. Although the setting of a three-person room doesn't align with the style I desire, I hope for a comfortable, slightly affluent, and bright living room. Perhaps introducing some chaotic elements to let the AI choose a more tragic setting.
The following is the resource list homepage of the resource station. It's interesting to see more well-known companies like Facebook and OpenAI updating here.
Comments
Post a Comment