At the request of one of the users of my models, I created a Ponyxl model in the style of Pilgrim_2D_SDXL. This is not a merge. One of the Ponyxl models was taken, only the blocks were cut out of it, and instead a data set was added on which Pilgrim_2D_SDXL was trained.
The result is not quite what I expected yet, but in the future I will still work on the model.
If you are interested in this model, do not hesitate to write comments, leave your works and like.
Description
Made based on the full version of Pony.
FAQ
Comments (7)
v1.0_full seems to have way better prompt / tag adherence than the previous version.
Also works really well with pony loras.
The style changed a bit, but it's still really nice. And it works without any style lora, whereas pony seems to need one to produce nice results (that or I am very bad at prompting for pony lol)
Thanks for your feedback.
I have a few more ideas to improve this model. If there are any tips that work differently in my model than in other ponies, write examples, it will help.
Well, for the most part I could create my DnD party very nicely. The only thing that is giving me a bit of a problem is trying to use ipadapter-face or instantid to make them look like the real life players. I did not have so much trouble with pony but I'm still trying things. I have absolutely no idea why this happens.
Apart from this, the model is great, I love it.
After further testing, it turns out pony doesn't work well with instantid o ipadapter-face either, so the issue sems to come from pony directly.
On the other hand, your v1 model works great and you can turn people into illustrations really nicely, so whatever you did differently between v1 and v1_full seems to be what makes it work or not.
On a similar note, v1 seems to work better with base sdxl loras and v1_full with pony ones.
I hope any of this helps you in your understanding of the models, I'll keep both and maybe try your other ones :)
@rootD Thanks for the tip. I can roughly imagine why this happens with the first version. It implements datasets and tokens from my large SDXL set, including a dictionary with weights. But it’s impossible to implement a dictionary with weights into a full pony; it simply stops generating pictures. Need to experiment with dictionaries...
@devpilgrim Well, whatever you did in v2 seems to have improved instantId by a lot. It went from producing very weird faces with artifacts in v1_full to making normal faces with a resemblance of the reference. It's not as good as v1 (because v1 is extremely good in this regard) but it's usable and with a bit of tinkering I believe it can produce very good results.
Also, v2 works well with both sdxl and pony loras, although their influence seems a bit lower than expected (could be because i'm mainly using style loras and your model has a very strong style). It's really a non-issue for me since a really like the style :)
Thanks for sharing your work!
@rootD Everything is connected here. It gets better in one place, worse in another. Finding a balance is quite difficult. As for loras, they actually perform worse, this is because I use the construction of sets of vocabularies and data that are usually used in standard style lores. It turns out that loras simply does not have an influence on the weight of words in dictionaries, as in simpler models.
Details
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.