Inference defaults fails to save settings / all settings disappear

[anchillabc]

What happened?

When saving inference defaults for a model, I expect it to be saved and remain saved. In reality, sometimes the UI works, saving the settings, other times, the settings do not get saved at all and the entire configuration simply vanishes when leaving the model page or the checkpoint manager. Worse, at this point, all of my previously saved inference defaults for all models have disappeared.

Steps to reproduce

• Open the checkpoint manager
• Choose a checkpoint model
• Activate "Inference defaults" and enter desired settings for a combination of sampler, scheduler, steps, cfg and dimension
• Alternatively to the above, enter desired settings first and enable "Inference defaults" afterwards.
• Leave the model page by clicking one of Next / Previous model, the checkpoint manager button or another button on the navigation bar, or the Civit AI link inside the model page (all tested)
• Switch to inference, using txt2img, choose model. Instead of just entered settings, the model comes with the default settings selected.
• Re-enter the model page; settings have disappeared, inference defaults is now turned off.
• Check all models, observe that all inference defaults settings are lost.

Relevant logs

Version

2.15.4

What Operating System are you using?

Windows

[mohnjiles]

The Inference Defaults on the model details page is only saving when you actually click Download on a model from that page. If you want to edit the Inference Defaults for an existing model, you'd right-click the model on the Checkpoint Manager page -> Edit Metadata -> scroll to Inference Defaults section -> make edits & save.

I understand this is not the most intuitive so we're working on making it work how you expect as well 😅

[anchillabc]

Not only is it not intuitive, but considering the model page actually usually contains recommended settings for a model, using the metadata editing screen would force users to go back and forth between the two screens to be able to enter all recommended details. If technically feasible, it may be a good idea, while looking at this, to consider adding a few more things to the possible settings for inference defaults, like VAE, refiner, and clip skip settings, or their explicit lack, to make switching between models easier.