Good job
It's a good start and I'm excited to see what the final model will look like. But right now I find it is extremely similar to pony v6, which is not a bad thing since it's still the best fundation model, but I don't see improvements over this 2 years old model. Instead I find lots of niche concepts existing in pony v6 but not in Anima. Can you help me make it clearer about what are the improvements?
Natural language prompting is such a massive improvement. Trust me when I say after we get higher resolution training this model will be better than ponyv6. I will say the multiple years of finetunes and training on ponyv6 have helped a lot with concepts. This'll have to catch up.
Now I'm really hyped ! Right now the natural langage prompting doesn't work very well for me. I don't know if I'm just bad at writing prompts, or if I tried too specific concepts
wtf? if you think this model is anyway similar to pony you have serious skill issues
Being french doesn't help :'(
My bad, I didn't test it properly. It's damn good! But it still has major flaws
- The hands are way better than pony
- some body parts are worse (I won't say which ones, but they are specific to males...)
- it follows the natural prompts much much better than pony, but it's still not very impressive because it has little awareness about the poses/size... for instance if I try to make anything assymetrical on someone's appearance, the model will ignore it.
- just like pony and diffusion models in generals, it struggles with clothes physics
My bad, I didn't test it properly. It's damn good! But it still has major flaws
- The hands are way better than pony
- some body parts are worse (I won't say which ones, but they are specific to males...)
- it follows the natural prompts much much better than pony, but it's still not very impressive because it has little awareness about the poses/size... for instance if I try to make anything assymetrical on someone's appearance, the model will ignore it.
- just like pony and diffusion models in generals, it struggles with clothes physics
i understand you are coming from p*\ny here but you really need to test both models on bit complex things, heck this thing beats noob in many aspects just not trained on highres properly yet.
clothes physics will be a struggle since most artist doesnt draw it accurately either.
comparing with p*\ny is so funny given how broken pony is unless combined with six loras and is an overfit merge of 20 models....
limitation exists but the difference between them is night and day.
You missed a few other pretty significant bullet points:
- It can make properly bright/dark images
- Artists and some characters (such as Aua and Pitvinm/Houshou Marine and Gawr Gura) are not obfuscated
- It's not fried - not skewed towards producing noisy and/or washed out images
- It makes far sharper images
And most likely
- It did not have its training data jpegged
- The aesthetic captions were not screwed up
Various sets of these were also solved in anime models before or after pony, though none except Neta/Yume did everything, and Neta/Yume have their own unique set of (IMO pretty big) issues with prompt adherence and undertrainedness. And of course there are ways to somewhat but only somewhat work around all those bullet points (e.g. pure black/white image 1.0 denoise), but this has those issues properly fixed.
And this model is going to get even more training which will make it even better.
If you like Pony you should try NoobAI. Illustrious/NoobAI has replaced Pony like a year ago.
That said, Anima is better than NoobAI and Illustrious by a magnitude.