Posts

Zecora applying lotion.

@Zerowinger
Coco’s always cute!

@Scarlet Ribbon
I’d say your half right. People are pretty creative, most just do / choose to do NSFW stuff, or so it seems.
Your completely correct as far as trying to figure out what the model understands and doesn’t understand (I recently did a Christmas themed image and found out that the model I was using didn’t understand what horse reins were, had to use a NSFW at very low weights to get it to work).

@Background Pony #A3FD
I’d like to see that too, more interesting SFW type stuff. It’s actually what I’m slowly working on right now. Working around what the model knows and doesn’t etc is definitely a challenge. Been getting help from various discord servers here and there but I’ve only just started down this path.

@Background Pony #A3FD
Part of the problem is that most people aren’t very creative with their prompts. I try to have a fair amount of variety, but there’s a lot of actions that don’t translate to ponies very well, and finding that intersection of ‘the model understands it’ and ‘the prompter thinks to try it’ isn’t always clear.

What I’d like is more images that are not saucy at all. I like happy ponies in particular. The problem is that most of those tend to be simple images of a pony standing or doing something trivial. I want to see more interesting images. I want to see unique styles. I want to see funny comics. Seriously, I can give you some 2-3 panel comic ideas which are at least better than nothing.

Regarding the lack of comments, Derpi has more but it also has more users. Tanta is small. Images usually get like 10-50 upvotes. 2 comments on an image is already a lot.

@Admin
I’m just relaying an explanation of what that error should mean. I’ve not seen it personally.

@sammykun
@Scarlet Ribbon
Are you sure that was the message? There aren’t any images deleted with the “DNP” in the reason, and the DNP system should be turned off here…

What was the video?

@sammykun
That means you are not permitted to upload images from that artist. DNP = Do Not Post.

@Lord Waite
Honestly, if Chroma-HD could be fine-tuned on Ponies and had better style consistency, I think it could be amazing. I’ve made some pretty crazy shit with it even without LoRAs already, but the inconsistency combined with generally slower performance is a killer.

@Scarlet Ribbon
Yeah, that is pretty terrible.

I wish someone else would just finetune a model well on ponies, so it knows the styles and the major and more minor characters…

(Quickly got tired of trying out pony v7, because so many other better models were coming out.)

Allow me to share a comment I wrote on PonyV7 discussion on reddit: https://www.reddit.com/r/StableDiffusion/comments/1oxd627/pony_v7_style_cluster_comparison/np04kc8/

TL;DR the style clusters have prompt bleed into each other, which makes the whole concept absolutely busted in a bad way.

@truekry
I’m rather shocked that one wasn’t on my faves list on Derpi and had its vote already migrated over. That is classic clop.

You can get more horse like features very easy. Like here.
I didn’t add “My Little Pony” to the prompt and tpony, and offshoot of ponyv6, made this from that.

@truekry
To your point, all my biggest peeves of AI ponies are the same as the popular trends I despise in hand-drawn ponies. I wonder how img2img using IRL horse photos as reference would change things…

[I have an M1 Max with 64GB of RAM, so I’ve got the power to run it locally. Just need to actually get things installed.]

@🐴
The thing is, we have little choice in that regard. The loras for that are all “big mode.” Getting smaller, oval shaped ones, aka more equine in nature, was always a battle fought with the models and loras (for the ponies in general) with a lot of manual editing.

But see, this feedback alone helped. We can talk about issues like that and why things are as they are.

@truekry
Some of it is that listing specific likes and dislikes would get repetitive and spammy for me. There are only so many times I can complain about the giant floppy crotchboobs everyone else here enjoys before I get myself worked up and become the first person to be banned for heterophobia. Likewise, “finally, a horse-shaped horse” has too little variety to be intellectually stimulating.

@Heat Sink
I leave a comment from time to time too, positive or negative. So far only 1 or 2 creators showed that they reacted and 1 lashed out at me for pointing out a mistake. It is what it is. I rather have it that way than total silence.

@truekry
About a month ago i personally did comment a bit on images that were posted here, but i only pointed out the positive stuff of what i liked on a picture. Mostly because i am unsure if the creator of said image would appreciate feedback on their images.

@truekry
I agree on missing feedback. Being told something as simple as “this is a cool idea” or “I like how this came out” would be nice.

@truekry
Hear here!

The worst thing about the silence is the missing feedback in my opinion. What did the people like about a picture, what did they dislike? Why did you get a down vote, why and up vote? It hard to guess. The only thing you can do for sure most people will like is plot to the face. I know, sex sells and all that, but I find it sad that we have this powerful tool that is AI, but if you want people to notice your art it better have plot in it. It also gives tantabus the vibe of a AI pony porn site and less AI pony art site.

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