Frequently asked questions and other tips (setup)

This list provides solutions to problems that you may have when hosting AIYA. I'll try to keep it updated with what I can think of.

highres fix does nothing

AIYA uses Web UI's legacy highres fix only. To enable this, go to Web UI settings and under the compatibility section, checkmark For hires fix, use width/height sliders to set final resolution rather than first pass. The legacy highres fix works by generating an image at a lower resolution, then upscaling it to the user's desired resolution using half the step count.

/identify is not showing the desired results

On your Web UI, head to settings and review the interrogate options. I'll explain the settings I know about:

• Interrogate: num_beams for BLIP - Think about the interrogate output like a tree with many branches. It starts at the root/trunk which contains the most obvious image caption; this is 1 num_beam. Each increase in the num_beam value goes one layer further into the branches to find possibly more accurate descriptors for the caption. Experiment with this value to see feels best.
• Interrogate: minimum description length - This controls how short your image caption can be. Setting it higher can lead to a more accurate caption, but also has the risk of duplicate words (e.g. "a cat with a bow on its head and a bow on its head")
• Interrogate: maximum description length - This is the maximum length a caption can be. Limiting this can help with the issue above. It's also advisable to limit length as Discord has a 1024 character restriction on the output field.
• CLIP: skip inquire categories - These are extra descriptions that get added on top of the standard caption. The AI seems to try to make a best guess on what categories match the image, but in my experience it's not great at it so I skip all categories. Your experience may differ.
• Interrogate: deepbooru score threshold - This affects the "Tags" phrasing option. I believe settings means the higher the value, the more confident the AI needs to be in the tag before displaying it.

"Tags" showing random unrelated tags - This seems to happen if the image is too small, such as trying to identify Discord emotes. Upscaling the image first improves the tagging accuracy.

LoRAs are not populating

The LoRAs do not yet have official API support, so I use a workaround to pull them from Web UI. It's a little inconvenient due to this.

Make sure your Web UI is at least on commit b7d2af8c7fa48d6eef7517a6fbc63a3507c638d4. This commit adds the LoRAs to the settings.
Then at Web UI settings, under Extra Networks, you should have dropdowns for the Hypernets and LoRAs. Click the 🔁 next to the dropdown list and it should refresh the list. Then click Apply settings, then Reload UI. Now on AIYA, you can run the /settings command, then Refresh. It should properly pull the list of LoRAs (and other lists).

changes to max_size does nothing

Discord is a bit resistant to sync some bot changes. If your changes are not updating in Discord, try using or opening the /draw command in a different channel. This short video demonstrates the issue and the solution.
This solution generally works for any time slash command stuff is not updating.

I don't understand the batch format

The batch field is a single value that controls both the batch count and batch size. It is explained more thoroughly here in the /draw command documentation.