Pony Test 2
This is a test on a LORA and art style for @d970685107384 . His request based of a LORA sent in a comment. His request was of Loita from Dark Hero Party. Upon a quick google search I found 3 maybe 4 images of her total. This would not be enough to properly train LORA to do whatever you wish, but to get a one off, simple gesture, sure It worked. I was able to get her to wave in a few models, but not much past the only main image on google.
After the first 6 models I tested, I tried the same prompt outside the trigger word for the style lora. Looking at the main image of Lotia on goggle she is very much a half-render/cell shaded anime style. A few tests on cell shaded I couldn't get it to simplify the details most versions so I looked to a Anime Screen Cap LORA to flatten the image. Anime doesn't have a crazy amount of detail frame to frame, so I needed less details.
The issue with his request was this isn't a popular character by any means. So getting this to generate with out any LORA support would be impossible. Asking for something known like Goku sipping a cocktail in a tux would have been cake for most any model. Why? Goku has most likely been trained as a data set and so has Dragonball Z. Naturally achieving the look would have been no problem for the model.
This test was destined to fail from the start. There was no win condition for it. If asked to create 2B in a wedding dress at the alter on the Moon, I bet I could have gotten that. But this character with a 1 image LORA was not going to net landing that "Style". Then for the sake of "But what if..." I still had the option to get the LORA for YorHA Style LORAs to easily recreate the style.
My conclusion:
I believe the people that are complaining about V2 > V3/V4 are use to creating very flat characters. Flat in this case, Low detail. Where V3 and V4 are very much not that model. They are trained on FAR more detailed data, IF not realistic. I even used negative prompts to sort that out. This also further drives the point, what they may be wanting to create doesn't have proper LORA support. With the additive of the Anime Screencap style, I instantly got results much more like traditional anime styles.
By using two other models as placebo and a control that wasn't based on Goofy_Ai's models to show similar results, that it's not just a Perfect Pony V. "X" issue. It spreads to other models as well. These results aren't cherry picked, I just ran the prompt a few times on V4 to see if it would generate or fail, then full sent the test. What you see is based on static seed, prompt and the LORAs.
In the end, I believe if you have access to the proper LORAs and or training data you can achieve any look you want in V3 or V4. The models are flexible enough to adapt, but they need to know what style you are looking for. If i type in: Chung-Li, wearing Elvis outfit, Standing on beach, in Dragon Ball Z Style, it would probably make this. But the random style I was given, there is a 0.00001% chance that image was originally trained in the model, and pulling that from all the training data would be impossible, where as a style like Dragonball is widely know and popular enough to easily make it into basic checkpoint training data.
TL DR: niche stuff = Proper LORA Support.