Confused about data separation

[set-soft]

First: I really enjoy this tool, is like an AI lab, and made it fun. Thanks!

Now my doubts: I'm really confused about the way data is searched and stored by the tool. IMHO looks quite weak, which is surprising for such a complete tool. I suspect too much influence from the poorly designed Windows environment (please don't take it as an offense, unless you are a Microsoft employee ;-).

The --data-dir option is very good, it controls where models and extensions are stored. Models are the heavy weights here, so this is good. The configuration for the UI is also read/stored from/to this path, good again.

But I'm confused about other things like: config_states and outputs. Both of them relative to the script place. So I don't get it. Am I looking it wrong? IMHO:

• config_states should be closer to the configuration.
• outputs is where you store all the generated data, the script directory isn't even writable if the python script is installed in your system, not just executed from a repo. I don't see how to configure its base.

Another thing that I see as strange is the fact that the script looks for stuff in just one place. This is really uncommon, most tools looks for data in various places:

1. Somewhere defined by an environment variable, skipped if not defined
2. User owned home directory (using HOME/USER or some XDG* definition)
3. System path (i.e. /usr/share/TOOL on POSIX systems).
4. Some hardcoded last resort place

Looking for internal data that is read-only using the path to the script is good, but not other things.

Some other related details that confuses me:

• The launcher does a checkout of various repos ... But it manually checks them out, no submodules and not shallow checkout (IMHO the best option here). Giving the fact that the hashes for the exact points are hardcoded I even recommend just incorporating the needed files to the main repo. At least one of the repos contains unneeded images.
• There are an option and a dir for the CodeFormer model, but the default behavior is to download it inside the CodeFormer repo.
• Extensions doesn't follow the main program rules, i.e. Control Net looks for models inside the extension.

Is just me or this needs some adjusts?

I'm trying to create some easy to install docker image with Pytorch+ROCm and the webui. I understand the webui have mechanisms to update itself and to save "working points" (config_states), but I prefer something more simple to install. So I would like to have a good separation between mutable and immutable stuff.

[Sakura-Luna]

https://github.com/AUTOMATIC1111/stable-diffusion-webui/wiki/Command-Line-Arguments-and-Settings

This is just an open source project maintained by individuals, and many functions are implemented by contributors of different levels. It is difficult to compare independent projects maintained by the team.
Extensions are implemented by different authors and WebUI has no control over them.

[set-soft]

I fully understand the project can't control the extensions. But can:

1. Provide policies about what to do and how
2. Provide clear APIs to solve common problems

And I already took a look to the CLI arguments, this is why I found --data-dir and this why I think the option to set CodeFormer model path needs some effort to ensure CodeFormer doesn't download it in the repo dir.

As for the open source: I can provide PRs to "fix" various of the above mentioned "issues", but I must be sure the core developers also agree that they are issues. Otherwise they will be wasted time.

[Sakura-Luna]

In older versions, ControlNet placed the model in the same location as the WebUI, but now it's moved to its own directory, it's not clear why.