keeps failing the hash check when downloading the model, issue is with v2.

Would you be open to training on anima? (or providing the dataset so I can 🥺👉👈)

I tried using it but it would just fail. I tried noobAi, pony & illustrious models, each try failed. Edit: V1 works, V2 fails. This is all on site. also I just noticed, it has 0 uses so far. v2 needs checking

NICEE thank you I love it <3

So i've been playing around with V2 for the past couple of months, off and on a couple of times a week at least, and i wanted to share my thoughts on it after using it for that long, the various good and bad aspects that i've been able to see from my own exploration of it.

Unfortunately, it does appear to me that, strangely enough, V1 is generally superior to V2 in some, perhaps even most aspects. I was very much looking forward to this Lora when i saw it get uploaded in to early access, and its one of the reasons its taken me this long to write about it - I tried everything i could think of to make it work the way i wanted it to.

Now, I will admit i haven't fully explored all the new capabilities that V2 introduced that weren't explicitly trained for in V1, but i will say that those features, evidently, have leaked in to the general use of the Lora.

I'm not too technical when it comes to Loras but i believe the correct term is "Overfitting"?

Using a strength of 1, it very much leans heavily towards two aspects in particular.

1 is that it leans towards a very westernish cartoon style and so, despite mentions of it working on other checkpoints, ive found that obsessionillustious is still generally the best one as otherwise the style interferes, or otherwise changes the base appearance of most other checkpoints to the point that you might as well not be using said checkpoint.

2 is that despite a strength of 2, and emphasis brackets in the negative prompt, no amount of "(((((((((Between legs, from behind, looking between legs)))))" etc will stop the lora from throwing 80% of generations towards that direction. It seems so long as genitals are mentioned SOMEWHERE in the prompt, the lora reeeeeeaaally wants to do one specific pose, where its from behind, with the character looking through their legs backwards, with the genitals being prominently on display even if you just wanted them to be included, or you wanted to specify which ones.

If genitals are not mentioned, then the first problem remains of the style overfitting. V1 generally didn't diverge from the style of ObessionIllust all too much, but V2 its very noticeably not the same style that is the checkpoints default.

If you use a strength below 0.8 or so does solve some of this at least partially, but you also start losing a lot of the loras main aspects. Instead of "Ruby rose as a dragon" you get "A red dragon with silver eyes with no hair", most of the time. Or in other words, likely what you'd get without any Lora.

Now, its entirely possible im doing something wrong but i just wanted to share my impressions, and maybe you could provide some guidance on how to avoid these pitfalls? Perhaps im simply using the wrong keywords that are triggering something within the lora that wouldn't NORMALLY be triggered. Still, it does appear at least to me that V1 at the very least needs the least amount of fiddling to get right.

What i will say though is that combining V1 and V2 loras, with V2 at a relatively low strength, does give some interesting results. It does for example, make aquatic characters like fish, dolphins etc much more "dolphin shaped" than what V1 alone could generally do. Its still far from perfect, but it does give much more usable generations than what V1 was able to produce.