I sat down to get some work done and tried what I shall call the cook book approach to using Stable Diffusion. What I found is the motivation for this post.
First a bit of background:
I wanted to make a LoRA. I tried kohya_ss. I cloned it, ran it as per instruction. It did not work right out of the box. It threw errors. It seemed to have trouble interacting with the file system, ignored my GPU. I could go on.
I also tried Automatic1111, ComfyUI–not for LoRA, but just to try–all with similar results. All left me with the prospect of debugging someone else’s codebase.
Are others having similar experiences?
As it stands, I am looking to take another direction, which means breaking away from Python. It has been said that such a thing is too difficult. But it is also said not really so difficult.
If you are in an environment where A1111 WebUI or ComfyUI does not work, it is highly likely that your GPU is not a standard one or that you do not have enough VRAM. This is a hardware problem, so it is a problem that cannot be helped because the PC’s performance is insufficient.
If your computer is weak, you won’t be able to complete the calculations.
Here, you have two options: either buy a better PC, or rent a cloud service for free or for a fee. If you choose the latter, Hugging Face, Google Colab Free, Lightning, etc. are all good options.
There is nothing Nvidia within miles of me. M1 has 64 cores with 128GB shared memory. I do not think we have a hardware constraint. Also, I do not think the difficulties with the file system are hardware. If I were to venture a SWAG, I would cast my suspicions on the gradio package. The buttons that are supposed to open a file dialog do not function. If I paste in a path, the code cannot find it and throws errors. I can go on.
Wait a minute. That’s the cause, and there must be some way to avoid it.
The WebUI and many other programs won’t work properly unless with nVidia GPU, and if not nVidia, special settings are required. Let me do a quick search.