/vt/ - Virtual Youtubers

>>66026043
Sorry but that is probably the issue, because the model will always learn classes it already knows quicker than classes it doesn't know which leads to things it already knows getting baked before your custom character name tag can encompass all that information under it.
In short if you try to encompass hair in the character tag 'xbhj' but your training tags also contain 'white dress' the white dress will get baked into the lora before your 'xbhj' tag can train all the hairstyle, hair color, face, length, body size, etc whatever else was pruned.
The resulting lora, especially on a small dataset, will be inflexible as a result so just tag everything in my opinion. Tagging everything also has the added effect of being able to add it into negatives later on, as "ch anon" (sorry no idea how people call you) demonstrated when baking the Non Anon lora.
A1111 has lora tag text autocomplete nowadays so all you have to do is click a thumbnail and it gets auto-added hence tag-flation is not an issue anymore.
Just my two cents after training 81 loras of small chuubas in the past 2-3 months.