Finally, an alright Hetero FPOV model
Important:
All new information regarding new versions of the LoRA are in the "About This Version" tab on the right. Recommend you read that if you're downloading any version other than 1.0
Also I'm a lazy bastard, the v1.1 LoRA has a different file name than the images shown, refer to the "About this version" for more details on my laziness, sorry about that.
This is a model I've been working on for a few weeks. I've been trying to look for good FPOV (female POV) models for a while, especially for hetero content (don't worry, there's other types of FPOV as well incl: futa, yuri, solo, masturbation, fingering, etc.). Problem is, there isn't that much hetero, FPOV images out there, so unless you can find some collections of good images out there that are hetero, this is pretty good for now. There's about 550 images in the LoRA.
Here are the tags:
FPOV
female pov
You could always try some different tags as well, try hetero if you want hetero. If you want some more creative results, try messing with the CFG scale making it fairly low or fairly high, also try changing the checkpoint. Be creative, experiment. If you get some good results don't be afraid to share :).
Credits:
It seemed to work best for me with FoxyalNSFW, but please give me more recommendations for other checkpoints to use with my LoRA. Using inpainting also seems to give some really good results.
Liudinglin's AbyssOrangeMix 3 (AOM3) has been a good model to use with this LoRA. Although, there are some weird artifacts that are also generated when using the model but are pretty small and perfectly fine. Some images I tried to generate came out completely wrong, but those were in the minority and some amazing images can come out of using this. Highly recommend using the model.
The LoRA was made using a guide by holostrawberry. You can make your own LoRA using their guides
Also, go check out and try these wonderful people's LoRA's, I saw how it could be done using their models. Model 1 for more futa and sole FPOV stuff, and Model 2 for more yuri stuff.
To see which model you should use check out the V1.1 post images at the bottom (idk, where it went, but here's the original post, haven't been on this site for a while) where there will be some images between my top 3 checkpoints that I recommend using (for NSFW, idk about SFW if you're trying to use this for that stuff) (also click on the post's name when you click onto the image post as it will bring you to some more info I have with regards to my testing). It will have the names of the models and the images they outputted. The models tested were: AbyssOrangeMix 3 (AOM3), FoxyalNSFW, and AnyHentai. The models were good at different types of images, so check out that post to see which checkpoint would be best for your image generation. What's being tested:
Revealing Clothes
Revealing Clothes
Naked Solo
Sex Position (attempted cowgirl, try different settings next time)
Masturbation
Ask me if you want any further information on anything, I'll try to get back to you to the best of my knowledge, but don't be surprised if I leave this model with no further updates, this is already my third attempt. Any recommendations are greatly appreciated, as well as sharing of images you create here (Follow the TOS though).
Lastly, I would rather you use this for your own personal use rather than to sell, but at least tell people how you made this if you are sharing.
Description
Actual Changes:
New version using some of the images sent by @guy33. Should be at least a little better now, you can compare this between the first version to see which one you would more likely use. Same tags but now "1girl" and "breasts" are more prevalent because it seems I forgot to include those when initially tagging the images. Also, there's now a total of ~850~ images (~300 new images). Pretty sure i had about 4,000 steps in this one. This should be at least a little better with some different positions. I have been getting more consistent results using cowgirl positions. Also there should just be more images of non-sexual stuff from FPOV, like just images from FPOV of legs/thigh-highs and stuff. Again, feel free to experiment yourself and share some images, no matter how "bad" or whatever you think they look, they might inspire others in what they want in their images. Mine are fairly garbage compared to some others because I don't really want to be spending a long time crafting my prompts, and I usually let the AI do what it wants for the most part. OH, really importantly, if any of you have recommendations of what checkpoint to use with the LoRA, that'd be great because i still seem to be getting the best results with Foxya1111's checkpoint.
Some Tips From An Inexperienced Prompter:
Experimenting with some stuff it seems that in general lowering the CFG scale is more likely to give stuff that looks pretty good, this probably won't be the case if you are looking for something super specific, but for most people (incl. myself), a low CFG scale is good for creating good images. A lot of the images in the training have the 'blurry' tag, it might be good to experiment with this as including it in your prompt can probably add more variety to what you get with minimal impact on the visuals, especially stuff like 'blurry background'. Having this slight blur effect makes the image usually look more POV'y (that's a word, shhhh) because our eyes don't focus on everything and choose to focus on specific things, the blur probably also helps the imperfections be covered up as you can ignore them more easily (ignorance is bliss after all). IDK about 'taker pov' it probably works, and it might be good to include that tag if you want more sex stuff, but when i experiment with it, it seems to have minimal effect (still good if you keep it in though... probably.).
Let me know if I fucked something up:
If you find that this new version is stopping you from doing some of the types of images/prompts that the previous version could, let me know in the comments and i'll try to see if i can fix the issue. Also if you have places where there are good FPOV of other positions and stuff that you think would be good, i'll probably add those, preferably with at least 20-50 images so that the LoRA can properly train on it.
Extra goodies for the more advanced good boys and girls:
Oh yeah, and i've decided to include a Zip file that has all the epochs of the training (it's labelled 'training data' because it has to have a file type for some reason, but it should be called 'fpov 1.zip'), so if you want you can experiment with more or less training (epoch is the percent of the total steps, so epoch 5 is 50% trained, while epoch 9 is 90% trained (Based on what i've read)). This might get you some better results with less training, however the normal full 10th epoch is usually the best in my testing. But if you want to test this, you can use the 'xy,z' script and switch between all the epochs that way (check out the LoRA guide, near the bottom on how to do it).
Wait, did i lie to you?:
So in the images that i posted it'll have a slightly different LoRA prompt (it won't say v1.1), that's cause i renamed the LoRA file AFTER i finished making all those images so don't worry, i'm not trying to get you guys to download something that doesn't work, it's just that i can't think ahead. Basically, don't copy and paste the prompts from my images and slap them in thinking it will work, make sure that you select the LoRA that you downloaded, otherwise it won't work.
This doesn't apply to those who use the "training Data" zip with all the Epochs because i didn't change the name of those before uploading (again, i'm an idiot who can't think ahead, and is too lazy to go back and fix it)
TL;DR
The LoRA got some more images, should be better at different stuff, and there are some tips that i would recommend using unless you are an experienced Prompter. If you are experienced, check out my extra goodies.
Anyways, thanks for checking this out, sorry that it's not in the description of the File itself, but here is my makeup present, happy prompting.
FAQ
Comments (7)
Even if it's not the focus, v1 is actually one of the better options I've found for 'receiving' pov with two female characters. Though getting that rather than the other way around takes some luck.
Yeah, i tried not to over-train my model on futa stuff. I tried to minimise the futa stuff because i personally like hetero stuff and if i trained it with a lot of futa stuff it might've tried to insert the 2girl stuff into the hetero stuff (you can already see that in some of the images i posted, some of the "guys" look feminine and have some feminine features like small breasts, mostly able to be avoided with the 'muscular' tag). Sorry for that, it seems that the checkpoints were probably trained using some POV stuff, but mostly from male POV, so that is probably leaking through and affecting how the AI tries to do it. Here's a post of someone who managed to get at least one image of what you're describing using a different model, also check out ChaoticsoulX51's post in V1 which has exactly what you're describing, you can try taking inspiration from them. You could try combining both of these LoRAs or even try inpainting/img2img. Don't get your hopes too high, but i might make an additional version for this taker POV futa stuff, or just girl on futa stuff in general soonish if enough people want it and i have the time; there's almost definitely enough images to train this so that probably won't be the hard bit. In the meantime, you can try increasing the weight of futa, taker pov
(e.g "(taker pov:1.3)" will increase the weight to 130% instead of just 100% (also think it's based on where the prompt is written, closer to beginning has higher weight, farther has lower.))
, and other stuff to try and course-correct the AI, as well as doing inpainting to try and switch the directions of who is taking it, if the only part that you don't like is who is taking.
Apart from 'sex' situations like in the examples, this LORA could be used for female POV in generale?
For example: riding a bicycle, stomping a door, etc ?
I think that you might be able to do that sort of thing, but it would be pretty hard without the correct checkpoints for the LoRA to use. I don't think i have those checkpoints and wouldn't really know which those are; additionally even if you did have the 'correct' checkpoints the LoRA is primarily trained on NSFW images and less so on common activities that happen to be FPOV. You could certainly try to what you're trying to do, but the LoRA is primarily for NSFW works, with occasional SFW images (you can see in my images that some are SFW but do have revealing parts to them.). Good luck.
I'm getting exclusively male pov results even when using the same checkpoint and copying the sample image settings directly. What could I be doing wrong?
A few things could be going wrong. Disclaimer, i can't see exactly what you are doing, so unless i see the images and exactly what you are entering there might be something going wrong.
1. Are you copying just the settings and not the prompts? Try copying the exact prompts. It could be that you have forgotten (or incorrectly) to apply the LoRA to the prompt as in the beginning of my images (insert the prompt using the little photo thing below the generate button and selecting LoRA and the current LoRA)
2. You copied Everything directly, even the LoRA in the prompt. When i was uploading the model i changed the file name but did not update that in my images and so it will not know what you are trying to reference. Insert the LoRA yourself and try again (use the method i said above). I said this in the "About this Version" (below the reviews), you can see that ChaoticSoulX51 used the proper LoRA prompt for v1.1 in their images which I should be clear in my post.
3. You are not using the tags given. Use the tags i gave, best used at the start of the prompt. 'POV' by itself will usually just default to male pov because that's what's more common.
Good luck and ask again if these didn't help.
Yo, it's been a while, and I haven't checked this site in a while either. Been doing a lot of stuff irl, and also been getting into local text LLM stuff as well. Those 2 reasons is why nothing's really happened here. If anyone wants the full actual training data, dm me on discord or something and maybe I'll see it. Send a dm/friend request to wailmer_fan and I might respond there, or if enough people ask, I'll see if I can set up a mega or something else type of link for the images. I'm in AEST so don't worry if I don't respond, and also that account is my alt, that I occasionally check.