You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
I have high suspision negative descriptors are considered as map descriptors. I gave the model a beatmap with just kiai and timing points.
Relevant settings used:
difficulty: 5
descriptor_1: clean
negative_descriptor_1: 2B
negative_descriptor_2: old-style revival
negative_descriptor_3: geometric
cfg_scale: 2.5
Used these twice, and both maps had 2B, old-style, and geometric elements, with star rating being 8+.
Tried putting cfg_scale higher and lower afterwards just to test whether that had any effect on it, and it seems like the higher it is, the more closely it follows the negative descriptors as if they are map descriptors.
The text was updated successfully, but these errors were encountered:
I did some testing and it doesnt seem like it uses the negative descriptors as map descriptors. cfg_scale with negative descriptors is just broken entirely and doesn't give the expected results. I got a bunch of maps that are just completely emptry.
I've had very similar behavior (following negative descriptors very closely), along with maps being cut short on high cfg (even without negative descriptors)
Try not using the 2B or Aspire descriptors and keep the number of negative descriptors equal to the number of descriptors, and see if the problem still occurs.
I used tech, slider tech and slider only as negative descriptors and jump aim, sharp aim and wide aim as descriptors with cfg 1.5. First 2 seconds were ok but then it got crazy slider heavy. I could send the mp3 file along with the rest of the settings and seed if required.
I have high suspision negative descriptors are considered as map descriptors. I gave the model a beatmap with just kiai and timing points.
Relevant settings used:
Used these twice, and both maps had 2B, old-style, and geometric elements, with star rating being 8+.
Tried putting cfg_scale higher and lower afterwards just to test whether that had any effect on it, and it seems like the higher it is, the more closely it follows the negative descriptors as if they are map descriptors.
The text was updated successfully, but these errors were encountered: