Posts

I don’t mind the natural language prompting as much, since I’ve gotten used to Chroma, but the style clusters seem to both be necessary and a total mess and it’s pretty unreliable. I’d personally rather have a version of Chroma fine-tuned on ponies, or something similar.

Someone finetuning ponyv6 or something illustrious/noob based so that 90% of the pony character loras weren’t needed would be great as well.

(Another model that has some promise is NetaYume Lumina, but, IIRC, ponies didn’t come out very well on it when I tried it.)

Hopefully pony v8 being based on Qwen will end up better…

Well, from what I remember, there was going to be version 6.9 based on SDXL, but it didn’t happen. Using AuraFlow as a base seemed like the best option at the time, since there was no real alternative like Flux Schnell or Qwen then.

Putting issues with development and deployment aside, the current PonyV7 version seems like a fundamentally broken model hampered by the initial decisions and the lack of proper documentation. Focusing on the natural language prompting instead of tags made the model a chore to use, requiring a ridiculously long and flowery prompt to get any useful results. Using style clusters instead of artist tags looked like an interesting idea, but it turned out to be an utter mess, considering no documentation on these clusters and them clearly not working as intended.

The end result is a model that is slow, inconsistent, difficult to use and creates outputs that are SDXL-level at best and SD 1.5 level most of the time. I don’t see any reason to use it over the current Noob-based, or even PonyV6-based models, which are better in pretty much every single way.

I’m curious to know what this community thinks of ponyV7.

Personally, I think it’s a good model, but it should have used SDXL as a base.. Training loRAs is impossible (for me atleast, as someone who uses a AMD 16gb GPU on windows + has 32gb RAM), AuraFlow isn’t… the greatest base model… if they NEEDED to use a model that isn’t SDXL, then they could’ve used FLUX.. etc

In short, they probably should have used SDXL as the base model, but that’s my personal opinion.

What do y’all think? :o