Sign In

A Study in CFG weights across models and with an "action" shot

00023-68748264-xl6HEPHAISTOSSD10XLSFW_v21BakedVAEFP16Fix-11.0.jpg
00002-2799864889-copaxCuteXLSDXL10_v2-11.0.jpg
00001-3709365836-copaxCuteXLSDXL10_v2-7.0.jpg
00003-2695548149-copaxTimelessxlSDXL1_v42-3.0.jpg
00005-628645591-copaxTimelessxlSDXL1_v42-11.0.jpg
00014-2641849954-electricDreams_electricDreamsV01-11.0.jpg
00018-2482561129-sdvn7Realartxl_beta2-3.0.jpg
00019-452680969-sdvn7Realartxl_beta2-7.0.jpg
xyz_grid-0000-1871435698-xl6HEPHAISTOSSD10XLSFW_v21BakedVAEFP16Fix-3.0-[generation_number].png
Images hidden due to mature content settings

Some thoughts:

  1. Wow most of these checkpoints are not good at people who are upside down. Or just this sort of "falling action" shot. The frequency of problematic arms and legs on this SD XL X/Y/Z run was very high.

  2. It's clear that I overstressed the models by combining "action" and "upside down." The only one that I felt even adequate including looks awful when rotated 180 degrees.

  3. CFG weight didn't impact basic anatomy error frequency. I didn't get the same level of detail-fixation as I got with my previous article and the genre accepts high saturation with much less difficulty than "foggy fantasy forest."

  4. Some checkpoints had 0/3 shareable images -- they just couldn't handle the scene's movement.

  5. I did notice that CFG 11 almost uniformly had the largest expressions of glee on the character's face. (Except copaxTimless). It is interesting that that was a "salient" feature that the models fixated upon. In this instance, correct as it communicate part of my intentional emotion of the scene, but... fascinating.

    Model creators -- let me know if you want me to make an "errors and rejects" post. Many of the models "best" images (in terms of composition and background) are ruined by bad anatomy. It might be useful to have a prompt like this as part of your QA suite.