Hassaku v2.2 vs Hassaku v3.0
I want to compare the last two versions of Hassaku.
My first brush with Hassaku was with v2.2 just some days ago, I find it incredible in generating incredible backgrounds, not to mention the way it generates character have a distinct touch when compared to, for example, WAI.
My only problem with Hassaku v2.2 is that it often generates six fingers, I was hoping v3.0 would fix that. In this post, I'm posting three sets of images, each set is a pair of images being generated with same parameters and prompts with Hassaku v2.2 and Hassaku v3.0, in that particular order.
The first set depicts an image of a male character who eats a sandwich on a food stall, with crowds that even included robots (I do prompt robots on this image) on the background and the foreground. Sure, Hassaku v2.2 generated a diner instead, but I am very particular with this gen because it really gave a sense of space (foreground, subject, background) however flawed it is in terms of details, but it ends up giving a sense of style.
That's Hassaku v2.2, now for the v3.0, with the exact same prompt and parameters... I was hoping to at least see the same generation, but, lo and behold, it changes dramatically. There's no robot in sight, there's one other guy for some reason, and the crowd is way too similar to other models. It is also visible in the image that the brown coloring theme is way toned down... But I do agree it is way sharper in terms of body details of a character.
The second set depicts an image of the same male character who is sitting on a hangar of dirigibles. Look at the incredible detail of the hangar, and the character, however imperfect it is being genned, looks way more natural in sitting down on that box.
That's Hassaku v2.2... Now, for the v3.0.
...I don't know how, but v3.0 really removes the details of the hangar, and the character looks way stiffer here. The brown coloring theme is absolutely toned down even when compared to the first set, and why is there a cat there? In terms of adherence, I do only put one dirigible and not two, but still... It also changed the robot to a sniper rifle.
The third set, and the last one, depicts an image where the same male character is shading his eyes while looking at some kind of mech in an open airfield, with the character's back facing the viewer.
I am very sad when checking this set in particular. The mech ended up being a filler in the background instead of the main focus of the image.
My take in this is that... Hassaku v3.0 is improved in character details, especially the hands, but it is achieved by sacrificing its style. I'm really dumbfounded that it changed a lot of things, especially with the extra character on the first set, the inclusion of a cat on the second set, and the change of focus on the third set.
I'm sticking to Hassaku v2.2 because its sense of style. I do hope the next versions would maintain the quality of character details while also keeping its style.
Thanks for reading.