Civitai Quick Tip: Importing from trainer and uploading a model have different default settings.

I reported this as a bug ages ago and was told they'd look at it, but still there so I figure I'd write an article on this oddity.

What differs?

Name/Type/Category (and Base Model)

This one is sensible: When importing from trainer the model will default to the values entered during training for fields it will know (though "concept" is so broad you'll need to change that most of the time). Defaulting new uploads to "Checkpoint" (over LoRA or null requiring the user select something) has always been an odd choice and one I see many new creators set wrongly, but that's a separate issue.

Default Permissions

This is a more impactful one. "Sell generated content", "Use on Civitai generation service", "Use on other generation services" default to unchecked when importing from trainer, but are checked by default when uploading a new model. The second and third only matter if "Sell this model or merges" is unchecked because it overwrites them, but it's an odd choice, especially as unchecking that is the most likely change anyone would make to permissions.

Trigger Words

This defaults to off when importing, unless you used the trigger word section of the trainer, then it defaults to having trigger words while listing only the one you entered. This is suboptimal because for most things you'll have supporting tags to include there, and (because the "trigger word" field just adds an otherwise normal tag to the front of the tag list) when you reuse a dataset in the trainer or upload a tagged dataset that trigger isn't distinguished from a normal tag.

Require users to be logged in to download this asset

For some reason this is disabled by default when importing, but enabled by default when uploading (including a new version. The on-site trainer still doesn't support directly importing as a new version)

Will they ever fix any of this?

Likely not. The glaring typo in the generator's "Content Restricted" warning is still there months after the bug report was acknowledged, and that one should be trivial to fix. Among the other issues with the trainer (my first article listed several that it still has) still says training LoRA variants is coming soon, and has said that since I started making LoRAs well over a year ago.

Pic unrelated.