BlackHat404

/

PooXL-PonyXL

like 11

First of all, thank you for showing enough interest in my work to comment. Using discussions is fine for this as far as I'm concerned.

Using this LORA together with watersports related tags ("urine", "peeing", "pissing", or "pee") will cause the poo to disappear and make the generated character spray pee out of their butt.
It could be a checkpoint difference issue but I after doing some quick testing about 80% of my gens include pee and poo when prompted. The placement of the pee is incorrect at times but it is rare enough(25% or less) that I would attribute that to bad seeds.

I'm going to be honest with you, I still haven't gotten a good feel for multi-concept LoRAs and that the fact that this one worked out still baffles me. Personally I'm not very interested in the cum in pussy part so I don't know if I'll add that intentionally to any future models but I might intentionally add peeing. As for FLUX, I'm not sure about training for it. As far as I can tell running it is at the edge of what my current hardware can do and I'm not super interested and buying time via runpod or something for this kind of stuff. In my experience FLUX does text really well but for me it takes about 4 times as long to generate anything and since I don't care about realism the quality difference that probably exists isn't worth it to me.

Sorry to be kinda negative but I don't want to mislead anyone about what will and won't happen. But again, thank you for expressing interest in my hobby.

First of all, thank you for showing enough interest in my work to comment. Using discussions is fine for this as far as I'm concerned.

And thank you for the quick response!

Using this LORA together with watersports related tags ("urine", "peeing", "pissing", or "pee") will cause the poo to disappear and make the generated character spray pee out of their butt.
It could be a checkpoint difference issue but I after doing some quick testing about 80% of my gens include pee and poo when prompted. The placement of the pee is incorrect at times but it is rare enough(25% or less) that I would attribute that to bad seeds.

Yes, I've had successful prompts as well, but this bothers me because anatomically, the pee stream should always spawn a bit away from the pussy because the urethra is closer to the front of the crotch, in between the clit and the vaginal opening. Theoretically, if a prompt contains "cum in pussy" along with "urine" and "feces" or related tags, the piss or poo wouldn't occupy the same space as the cum, but most checkpoints can't differentiate the groin area as 3 separate compartments with a front, center, and back unless it's particularly specified in the training data. I think this is also why sometimes if I prompt "feces" and "pooping" with this LORA on an image where the character is facing the front, the poo would come out of the vagina. It's also a bit hard to get accurate results when prompting POV penetration if the character is also peeing.

I think it could be solved if the training captions can be made to teach the model how to differentiate the groin into regions where only pee will spawn, only cum (or a penetrating dick) will spawn, and only poo will spawn. I'm using the example of "cum in pussy" more as a landmark to train in the location of the vaginal opening relative to the urethra and anus. In hindsight this is only a problem for me because my prompting skills might not be good enough for my complex ideas and I'm a stickler for not having to reword my prompt over and over to get the result I expect.

I'm going to be honest with you, I still haven't gotten a good feel for multi-concept LoRAs and that the fact that this one worked out still baffles me. Personally I'm not very interested in the cum in pussy part so I don't know if I'll add that intentionally to any future models but I might intentionally add peeing. As for FLUX, I'm not sure about training for it. As far as I can tell running it is at the edge of what my current hardware can do and I'm not super interested and buying time via runpod or something for this kind of stuff. In my experience FLUX does text really well but for me it takes about 4 times as long to generate anything and since I don't care about realism the quality difference that probably exists isn't worth it to me.

Understandable. And glad to hear that! Again, I think the results could be better if the model can somehow be trained to separate different regions of the groin.

Yeah FLUX is still really slow for me too, but the fact that it can get hands and text right without complex prompting, and the fact that it understands natural language way better than SDXL checkpoints so I don't have to memorize the tag lists of 3 different boorus is a godsend. Also realism is nice, but it's more the convenience for me.

Sorry to be kinda negative but I don't want to mislead anyone about what will and won't happen. But again, thank you for expressing interest in my hobby.

No problem! Honestly I'm just thankful this exists in the first place. lol

I will back this request as well. Having a hard time getting pee from vagina and poo from anus.

the fact that it understands natural language way better than SDXL checkpoints so I don't have to memorize the tag lists of 3 different boorus is a godsend.

A lack of proper booru tags is actually one reason I don't like base FLUX lol. I'm used to booru tags so "1girl, dark_skin,futanari, penis" feels better to me then "A dark-skinned woman with a penis".

I think it could be solved if the training captions can be made to teach the model how to differentiate the groin into regions where only pee will spawn, only cum (or a penetrating dick) will spawn, and only poo will spawn.

Interesting idea. Maybe tagging images with "pee" with "urethra" as well would help with this. I think that adding anus to most of the captions for my early scat based models helped a lot with placement. I'll look into it when I try the combined pee and poo model.

Hallo, I've had some luck.

The data set is smallish(about 100 total images 25 pee only, 25 poo only, and 50 with both) but conceptually I think it shows promise.

This does look promising!

Hallo, I've had some luck.

The data set is smallish(about 100 total images 25 pee only, 25 poo only, and 50 with both) but conceptually I think it shows promise.

Can it even work on Anime base?

Can it even work on Anime base?

It's trained on PonyXL so all Pony based Models should be compatible.

Any idea when we’ll get a chance to try this model?

Didn't really have time to run tests on it after posting. I might have something to post over the next few days depending on how testing goes.

Hallo, I've had some luck.
The data set is smallish(about 100 total images 25 pee only, 25 poo only, and 50 with both) but conceptually I think it shows promise.

Awesome. Looks like it's going well! I'll be watching with great interest.

Hello again I've made some progress and wouldn't mind some help with testing it before it's officially posted. Top row is model on and bottom row is model off. You can find the beta model here https://huggingface.co/BlackHat404/PooXL-PonyXL/resolve/main/ExcreV1.safetensors

It was trained on "omoko" but that doesn't seem to effect the output much so you can probably leave it out. Important terms are "peeing, pooping, urethra". From what I can currently tell, laying down isn't well supported but I kind of just threw the dataset for pooping together based on filesize so I'm not surprised.

I have also started a Discord to assist with any beta testing(if you want, no pressure) but I don't wish to post the link publicly so if you are interested in that. You can DM me on CIVITAI. Please don't spread the link around because I don't want to deal with a lot of moderation due to bots finding the link.

Happy to help. I should have some time this evening or tomorrow. I’ll DM you on Civit to link up on Discord.

Hello again I've made some progress and wouldn't mind some help with testing it before it's officially posted. Top row is model on and bottom row is model off. You can find the beta model here https://huggingface.co/BlackHat404/PooXL-PonyXL/resolve/main/ExcreV1.safetensors

It was trained on "omoko" but that doesn't seem to effect the output much so you can probably leave it out. Important terms are "peeing, pooping, urethra". From what I can currently tell, laying down isn't well supported but I kind of just threw the dataset for pooping together based on filesize so I'm not surprised.

I tried it out for a bit using a simple prompt and a few that are more ambitious. Grids are strength 1 vs strength 0 and with/without "omoko".

Results are more accurate with the LORA active! There are some slight differences when using "omoko" and from what I can see, it seems to help keep the urine streams consistent when the prompt is more complicated, or when you're using another LORA that could cause conflicts. Otherwise, I don't see too much of a difference in terms of quality unless I messed up using the options in the X/Y Grid settings. Also, it still has slight problems with "cum in pussy" or similar tags, but it's easier to get them to work together by just adjusting the prompt order and wording instead of straight up needing to inpaint.

And this is more of a side note, but from personal research in regards to models based on PDXL, most checkpoints can inherently recognize the tags "urine", "pee", "pissing", "peeing", "feces", "shit", "poop", and "pooping" without needing to use LORAs.

"urine" along with "feces" and "shit" are usually more powerful and can sometimes generate peeing and pooping scenes without actually adding "peeing" and "pooping" as tags. Also, "pissing" is more consistent than "peeing" but is more likely to make the stream come out at a weird angle for female characters, usually shooting straight backwards with the character facing away. Presumably this is because it's a derpibooru tag and a lot of the data it's trained on might have depicted uh... horse anatomy.

Anyway, keep it up! IMO this is a great improvement overall.

Thanks for the insights. I was aware that peeing and pooping were supported by base Pony but it wasn't as well supported as I would have liked lol. Overall it looks like this version is doing pretty well so the next version will probably be the release version.

@rangifertarandus77 Would you mind sharing some of the prompts you used? And maybe a sense of the checkpoints or additional LoRAs you your examples?

Thanks for the insights. I was aware that peeing and pooping were supported by base Pony but it wasn't as well supported as I would have liked lol. Overall it looks like this version is doing pretty well so the next version will probably be the release version.

Yes they're pretty weak on their own, but as I've said, from personal experience "urine", "feces", and "shit" are much stronger than both "peeing" and "pooping" in base Pony. I've scanned a few of your prompts and you only use "peeing" and "pooping" by default, so I recommend you try either adding them in or using them instead of "peeing" or "pooping" to try and compare the results.

EDIT: I meant adding them in when you prompt images using the finished models, not adding them into the training data. That would complicate things too much, I think.

Awesome news, can't wait!

@rangifertarandus77 Would you mind sharing some of the prompts you used? And maybe a sense of the checkpoints or additional LoRAs you your examples?

The metadata are intact in the grids I've posted. If you run SD locally you can drag and drop them into "PNG Info" on A1111 or Forge to check the prompts, or use a standalone tool that can read prompts from AI-generated images, like https://github.com/receyuki/stable-diffusion-prompt-reader or https://sdimage.info