Note #1 And VERY IMPORTANT: THIS IS SUPER BETA RIGHT NOW, PLEASE BE RESPECTFUL. You will likely only get ok images one out of every 25-30 tries, depending on complexity and following my initial prompts. I'm looking for feedback on what works well and what doesn't as this one is so fucking complicated and I was getting stuck :). Let me know how your tests go, just don't blame me if you get crazy monsters. Oh, but do have fun with it! And it felt like it worked best on AbyssOrange but I could just be delirious with that after staring at so many pictures.
Note #2: A second upload for today, namely cause I'm tired of working on this one for now and need a little break and feedback. I've made too many version to count :(.
What does it do you ask? Well, 90% - 95% of the time not what you want. The 5-10% it might work, it'll give you your waifu wearing whatever you designate in front of you, while the viewer is holding a camera with her in a different outfit and state of undress. Importantly, she should be more skimpy looking on the phone then in real life. Now, is this her sending you the pic? Her new lover? I dunno. Whatever you want to RP it as.
How to use it:
(I use Sampling Method DPM++ SDE Karras and Upscaler R-ESRGAN 4x+ Anime6B) (Let me know if something works better for this Locon)
Highly recommend not using hires.fix immediately. Make a batch of like 20-40 images, fine one you like, then use hires.fix on that one to try to fix it. I didn't use inpainting in any of these images but that will obviously help a ton as well.
I have found the following prompt to give me a decent amount of success, though it needs tweaking on each model and with what you add. Please try this AND try your own versions to see if you can get a better basic prompt working. I'll thank you forever. And just give me general feedback so I can make a V2.
<lora:YourLoraNameHere:0.9>, PhonePic, (viewer holding phone, phone screen), netorare, 2boys, penis, sex, vaginal, nude, nsfw, ((sundress)), 1girl, full body
Now, breaking it down:
<lora:YourLoraNameHere:0.9>, PhonePic, (viewer holding phone, phone screen) - I've been putting some version of this in every prompt.
netorare, 2boys, penis, sex, vaginal, nude, nsfw - Put here whatever you want her to be doing on the phone. Replace COMPLETELY with just an outfit if you just want her chilling in the phone pic.
((sundress)) - Put here whatever you want her to be wearing in real life.
1girl, full body - Not mandatory. I'm just trying stuff out.
Extra, Extra!: This won't work well. I warned you :).
Picture Descriptions:
1-3: AnythingV3 using just this Lora and my prompt style
4-5: AbyssOrange using just this Lora and my prompt style.
6: RevAnimated using just this Lora and my prompt style.
7: AnythingV3 combo'd with my in progress Vivy Lora.
8: AnythingV3 combo'd with my released Tohru Lora.
9: AnythingV3 combo'd with my in progress Knight-ify Locon.
How was it trained?
It was trained on animefull_final with variable amount of images with the usual self-pruning and tagging and a wide range of different images. This specific version has 25 images because I cut a lot of them because I think they were making it not work too often.
What's next for this?
Any sort of improvement I can squeeze out of this... Which is a lot.
What's next that isn't this?
More character Loras. (Viper and Rupee from Nikke, Spectre Tenebria from E7, Zephia and Yunaka from Fire Emblem Engage, Ivy from Soul Calibur 5, Heine from Combatants Will Be Dispatched, Conqueror Lilias from E7, Evelynn from LoL (both base and All Out version), Tharja from Fire Emblem, Laharl-Chan from Disgaea 5, Miku from Idaten Deities, Kuribayashi Shino from GATE, Rin Sohma from Fruits Basket, Kasumi from P5, Ann from P5, Ikumi Mito from Food Wars, Alice Nakiri from Food Wars, Chika Fujiwara from Kaguya-sama, Lulu from FF10, Momo from My Hero, Sheik from Zelda, Vivy from Vivy, Lucoa from Dragon Maid, Laharl-chan from Disgaea, Dominique de Sade from Vanitas, and Dos from One Punch Man.) And maybe Wilhelm van Astrea, my favorite Re:Zero husbando or a Prinny from Disgaea.
V2's of some of the previous Loras. Specifically make them more diversely acceptable with other models and LORA's.
I also have some more quite NSFW Locon's. I decided to list them as well, so if people have preferences on which to fix up first let me know. Specifically, Beta Censorship (Automatically censors any images you create to make them BetaSafe), PhonePic (Thisone has been very, very hard... But TLDR, someone holding a nuder version of the waifu on the phone), Demonic Transformation, being stepped on, Robotification, Defeated (decently NSFW defeat scenario for the waifu getting abused), Petrified (I'm unsure if I should bother with this one anymore), Infection (this one is VERY nsfw, as the waifu is completely infected), Dominatrix (waifu becoming a dominatrix), Brainwashing Machine, Tentacles, Pillory, Masked Evil (A mask that possess' the waifu), Latex Corruption, Combat Drone (hard to describe but basically making them a mindless sci-fi combat warrior), JAV Supervillain, Jiangshi (Transforms into a Jiangshi), Mind Flayer Defeat, Blacked, Hypnotized Harem, Hourglass Figure (just to make your waifu curvy), and Holding a Leash.
And, uh, I want to make a few "normal" Locon's so I have an "knight-ify" your character one coming up. And a forest fairy one as well. And a officer one. And a barbarian one.
Additionally, I love the style of Defaultz_17, so I'm making a Locon to mirror that.
Or just give me some random suggestion or priority opinions. That works too.
I've also added a couple of my extreme NSFW preview images to my twitter (https://twitter.com/TwoMoreTimes89). Let me know if you think I should work on them more for a submission here.
Description
FAQ
Comments (26)
This submission probably doesn't deserve any positive reviews. But I'd appreciate it if you don't drop any highly negative one's either, as I know this isn't fully working but would prefer to let people play with it and give me feedback.
If it seems to be attracting too much negative attention, I'll unpublish it temporarily till I get a better version up.
But please do give feedback and opinions on it and if I should continue working on this.
Seems like a very interesting and unique Loco/LoHA. Will for sure give it a try later tomorrow.
Great concept! Kudos for the effort. Would definitely like to see something like this working more consistently. But it seems very challenging to train.
I might have a go at training something similar later on, just to take on the challenge, and in the hopes of not hating myself for spending so many hours trying to squeeze something out with little success :)
Dude if you can get this working you will be the greatest thing to happen to humankind since sliced fire wheels were discovered. Super excited for this one
hope can twaek lora display name, for better download name, make it more readable
可以的話 希望能調整一下 你的 lora 名稱的顯示方式 不然下載後 都因為前面的描述字導致 你的 lora 檔名 全都沒有識別性
Yup, definitely will once it's not in Super Beta. Want people to pay attention when they download this one specifically :).
I don't have any images I really feel like sharing. This one isn't really up my alley anyway. In 40 images I only got 3 that were in line with what the goal was. Most results were:
-Girl holding the phone out for the viewer
-Phone positioned over the girl's body with a skimpy outfit on the screen, but blocking her actual body. Clearly the AI is trying to just generate one subject.
-Girl's body melds into the phone into a perspective nightmare.
I'm sure you've seen all of these.
I did try using an additional character LoRA. The non-phone girl was usually accurately depicted. The phone girl usually had similar features, but was also clearly not the same person.
Good luck.
Yup, the current hit rate is quite low and each of my submitted pictures were ones taken from maybe 30 productions each. Basically mass-batch 20-30 and would select 1 that was good enough (which is now part of the wheel I have up there). Appreciate the attempt though! This one will be in the slow burn with updates while I just continue with my normal production for other stuff.
It's called Photo_Comparison the actual tag
Hmmm, where did you find that? I tried hunting through danbooru and reverse tagging images and didn't see it. If I could find a larger sample size using that tag it would definitely help.
@TwoMoreTimes89 Actually I had created this Lora long ago, then Didn't post it because it wasn't giving proper results, but after Seeing this post I got courage to post it. You can read the description on my Profile. Also would you mind exchanging Datasets so we could produce better images.
I don't know, for me it's working kinda consistently
@lolikkrun Are you using the base prompt-style I've been trying or one of your own concoction? And which model (Abyss, Anything, something else?)
it seems to work 95-98% fine with my tries, make sure using LoCon add-on to trigger the model (<lyco:SUPERBETATESTINGChangeACharacter_superbetaversion:0.9>) not with LoRA. the biggest issue is the hand holding the phone...
I am unsure how to use the LoCon add-on and I can't find a straight answer online, could you point me to a decent resource to read on it?
@blackroseyagami Hey hey. I've begun to add some simple descriptions in my future Locon's that go into the basics for setting it up. Basically this -> "As this is a Lycoris/Locon file, you will need https://github.com/KohakuBlueleaf/a1111-sd-webui-lycoris extension to use the <Lyco> tag. If you want to use the prompts like I do, specifically being able to use <Lora> instead of <Lyco> (only effects ease of use), you will need this as well https://github.com/KohakuBlueleaf/a1111-sd-webui-locon. Either of these work, you will just need to choose between using <Lyco> and <Lora> prompts."
Let me know if that's enough, otherwise I can go into more detail!
Thank you mate!!!
I totally had the 2 extensions confused as the same one, that is why I wasn't able to get them to work.
I am now running experiments with it, let's see what happens.
@blackroseyagami Dope! Let me know if it doesn't work!
This one is really nice, it's finnicky, but the results are really something else.
I'd really like to see the defeated and tentacle locons as well, all of your work seems really interesting!
@crimelad I'm debating on just how NSFW I want the defeated one to be. I have an extreme one ready and a more mild version. Any preference on your end?
Definitely finicky. I didn't have much success but it's incredible that it's working as well as it is given how complicated the concept.
I've had some luck running some generations, and then modifying the content of the screen with a separate inpaint generation! Quite happy with the results! :D
Would you be willing to share the method used to make this locon? Very curious on how you tagged. As well as what type of images and how many were used. Trying to make something similar with x-ray.
@barbecue420 I can share the dataset stuff later, as I'm currently not at home for a work trip so don't have access to my main computer.
In the meantime, I did actually try out an x-ray version, specifically around someone holding a pair of glasses in front (like the phone) and the girl's clothes would be x-ray'ed only through the glasses. It was probably a little less consistent than this phone one, so I took a break. My biggest issue revolved around the glasses being glasses that actually looked like something someone would wear and not a deformed shape... And then also the girl being appropriately clothed.
If it's helpful for you trying, what I tried to get the medium success was keeping the type of x-ray consistent (glasses in this case), as I initially tried being flexible but it get's the Locon's confused. For the dataset itself, I erased tags involving glasses (and also no pictures with the girl wearing glasses as I worried it would confuse it) and I erased tags that involved the "x-ray'd" part of the girl (like midriff, or underboob, or thighs, etc). Then my prompts would be builg with the Locon activation, plus the location I wanted x-rayed. Aka, breast for see through tops, underboob + midriff for see-through lower breasts and belly, etc. Now was that the right way to do it? Hell if I know, I still needed to try more things lol. But at least that was my attempt and maybe you can get smarter/more creative with it.
Hopefully that was helpful and you can make a better version than my attempt ;9. LMK and @ me if you any other questions! I'd be excited to see a good x-ray version pop up on civit!
Really good idea, here are the issues I found:
- The image generated inside the phone is really low quality or is inhuman (merge with other things or have like way no many hands);
- The hand holding the phone is often bad (too many fingers or not enough many of them, the hand merges into the phone or photo in it, etc.);
- The position of the phone is often hiding the main subject or merging with it.
Those aren't really bad issue as they mostly can be fixed via photo editing, but going this way I think this would be even easier to just generate the hand and phone with a simple background like a green background. It would allow to generate 2 more images (background girl + photo on the phone) separately then just use a color key (for example) to remove the background then add up all the 3 images together.
TLDR - Good job for a "BETA" and works almost fine (at least to me) but I think it would be easier to only generate the hand and the phone then add everything together (personal opinion).
OMG!!! PLS DO A pony or SDXL version!!!!
Details
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.