Power Pack's older sister. This time (hopefully) trained when the trainer wasn't completely broken.
Head: G_H_J_P, red hair, long hair, blue eyes,
Outfit: two-tone bodysui, tred bodysuit, metal vambraces,metal boots
I unpublished the old, broken, version.
Description
Comments (12)
Looks better now! Was it trained with preview3 version of anima? Is that the default?
Default. Don't even know enough about Anima to know preview3.
@NanashiAnon Ok, the showcase pics say preview3, so I guess the civitai trainer is using that as default (and not the latest iteration anima-base_v1.0).
@RisingV I see. I guess someone who knows more about Anima should report that.
@NanashiAnon I can write a report, though I haven't used the civitai trainer, so I can't report first hand. But as much as I understand the images in the showcase are pics generated during training with the same checkpoint the LoRA was trained on, is that right?
Maybe using preview3 is intended behavior though...
@RisingV Hmm, then not sure what I'd report
"Metadata shows on-site trainer for Anima LoRAs is based on outdated preview3 version. No reason it shouldn't be using anima-base_v1.0." ?
@NanashiAnon Something like that. I just donwloaded ita and looked at the metadata and it's saying the model name is: urn:air:anima:repository:huggingface:circlestone-labs/[email protected]
So either this metadata is wrong and preview3 was used for training instead or the metadata of the showcase images is wrong or it was trained with base-v1.0 and the images were generated with preview3 instead for no reason. Either way there is something off.
@RisingV Just make sure: You're saying the square ones are claiming to be generated with preview3? The horizontal gallery images were made with Hassaku Anima checkpoint post training.
@NanashiAnon Yes exactly, you can see it in the metadata of the image and if you click on the model's name it's directing you to preview3 version. I also downloaded the image to check the metadata independently and it's saying "modelName":"Anima","modelVersionName":"preview3-base", so it is generated with preview3 (if image metadata is correct).
@RisingV Got a response
"Hi NanashiAnon,
Thanks for reporting this, and for the detail.
You're right that there's a base model mismatch affecting the Anima training samples, with the previews being generated off an obsolete base. We're aware of it and are looking into it.
Thanks again for flagging it."
@NanashiAnon Good you reported it then. If they were aware of this, why did they not change it on their own? Or should this be read like "Now we're aware of it."?
@RisingV I have no idea with the responses I get to bug reports half the time.