🎶 If you like it, you better put a Cock Ring on it! 🎶<hr />KREA2 version 3: I noticed that my previous version, v2, was interfering too much when stacked with character LoRAs so I created a brand new HQ dataset and trained it with a much improved workflow compared to what I was using when I created v2. It is now much more capable, flexible with prompting, and stacks much better. I included a large number of images in the dataset to allow maximum flexibility so just describe exactly what material, color, visibility and other factors in the prompt and you'll likely get it either in the first or second shot. KREA2 version 2: I'm really liking Krea2's prompt adherence and LoRA qualities are great with not a lot of steps required to learn. I'm still figuring out the optimal AI Toolkit settings that balance training speed and output quality, but my current approach yields some pretty good results. There's nothing fancy you need to do with this one - it's very straight forward. You can specify material type, thickness, numerous compositions and viewing angles and with a solid prompt you'll get what you want. Examples provided in image gallery.For more spicy prompts, you'll need to make sure you have a KREA2 filter bypass config - for example and details see <a target="_blank" rel="ugc" href="https://civarchive.com/models/2746817/krea2-filter-bypass-fedor">Krea2 Filter Bypass [Fedor] </a>or if using Comfy then an alternative you can use a Krea2 prompt conditioning enhancer / bypass node; there are several variants - I'm trying this <a target="_blank" rel="ugc" href="https://github.com/capitan01R/ComfyUI-Krea2T-Enhancer">one</a> out currently.Flux 2 Klein 9B v3: Dataset from v2 but ran through Flux 2 9B KV with prompting to add realistic skin texture and color correct / relight - results should be better on realistic images with less plastic-like or and reduce overly smooth skin artifacts.I created it primarily for edit image / inpainting usage but is fully capable of full image generation as well.Flux 2 Klein 9B v2: Mostly from ZIT v2 image dataset (with some improvements). Intended to be used mostly for inpainting / edit functions of Flux 2 Klein but it can easily do full image generations.Z-Image Turbo v2: Significantly improved the dataset with HQ images only, upscaled with SeedVR2 and removed any previous images that weren't exactly right with the cock ring mechanics which should help eliminate any ring blending into the skin issues (it may happen but it'll be a lot less). The dataset was varied enough that you pretty should be able to create them out of any material or style you want, so long as your prompting is sufficiently crafted. I even made glass/crystal ones out of this version although the dataset itself didn't contain any.I've also tested it with stacked character LoRAs and it seems to do well in those scenarios which is a bit surprising - I haven't done extensive testing on this yet, but at initial testing results are positive but your mileage may vary.A side benefit is that this can generate gentleman's vegetables reasonably well too, so a double pass with this and then a peen LoRA shouldn't be necessary.Z-Image Turbo v1: Pruned and upgraded dataset from my Flux versions, and has a great deal of flexibility (thanks Z-Image!) so all that matters here is how you prompt. During testing, any body horror or image strangeness could be corrected through refined prompting or a slightly different seed. All images in model gallery are unaltered, albeit cherry-picked, ComfyUI txt2img generations - with the obvious exception of the img2vid examples generated with WAN.Strength 0.5 - 0.8 for realistic, depending on your checkpoint 1.0 may create more plastic skin like characters.FLUX1-Kontext version: Made from approx 142 images, of different types and angles. Trained with 'add a [x] cockring] where [x] is the material. Materials expressly named are leather, metal, rubber and of course glowing (why not - it's AI image editing, so might as well throw in some surrealism).FLUX1-D version: Similar training approach with my other Flux Jockstrap LoRA, this with a dataset of approx 500+ HD images, using the same LLaVA prompt captioning parameters that focused on object + context approach. While its size a bit smaller, it's a result of the smaller dataset - I'm still on the look out for a better online trainer that allows me to specify more parameters other than the Replicate trainer I'm using so I can optimize LoRA outputs better.Like the last one, I used a cured dataset from previous Pony/SDXL versions, with the same low and slow training approach.This version works fairly well after I adjusted my captioning strategy. My other versions, not published, were hit or miss but since this one is fairly reliable and predicable, so I thought I'd share it here.I initially intended to create this as a strict inpainter but I noticed that along with a cock ring it can generate male genitals at an okay level as well. A likely result of having to specify peen in the captioning process even though that was not the focus of the object-in-context training captioning. So if you're using this to generate a whole image (hole image?) and it generates a subpar/nightmare peen then I recommend either a Flux model that knows male anatomy or probably the easiest would be add on a peen LoRA (this one should play well with others). I generally go with the latter approach. If that doesn't work for you then there's always the 2nd-pass inpaint route as a last resort.You can specify material, such as plastic, rubber, leather, metal, etc. As well as colors and since I included the previous dataset that had a bit of glowing cock rings in it - so you can even make'm glow. No batteries required! How 'bout dat!<hr />Pony/SDXL versions: It has multiple styles, such as leather, metal, plastic, etc and you can also specify colors. Basic trigger is pretty easy: cockring but then you can get creative like pink studded yellow leather cockring, or cobalt blue metal cockring , etc.<s>Made this pretty much an inpaint only LoRA in mind, so keep that in mind. You may be able to get good results if you generate the full image with it enabled but I found that if you generate what you want and then add it later it works much better that way. </s>I've found between 0.6 - 1.0 work well but mileage may vary.Made with the intention to be used as an inpaint LoRA, but does work for full image generation - but be mindful if stacked with a character based LoRA the fidelity will likely be affected.--Update: Pony v1--Why a Pony version when many base checkpoints know the concept already? I found most realistic Pony checkpoints can generate a cock ring on their own, they either don't a varied material or aren't able to handle varied perspectives. So I made this one.. also tossed in a glowing material type as well just to mix things up lol

Put a Cock Ring On It! (KREA2, F.2 Klein, ZIT, F Kontext, F.1 D, PDXL & SDXL)

cockring_krea2_v3.1.safetensors

cockring_krea2_v3

cockring_krea2_v2

cockring_F2K9B_v3

cockring_F2K9B_v2

cockring_ZIT_v2

cockring_ZIT_v1

cockring_FLUX_kontext_v1

cockring_FLUX_v5

cockring_PDXL_v1

cockring_SDXL_v3

cockring_SDXL_v1

Files with SHA256: aeafaecd8c7fc9b7b39f8a9392dcebee9cf6eb0795e425ee2b93d524c77a805a