FLUX 2 Klein 9B version 2: Retreated v1.2 dataset with a Flux 2 9B KV wash to get realistic skin textures and color grading / corrections. I kept the captioning the same, so flexibility will be largely the same as 1.2 but with better skin and fabric texture outputs.I use it mostly using Klein's image edit mode but it's versatile enough to full image generate what your probably looking for with a suitable prompt. It can go from plain Jane jocks from front, side and back-ass views (may take a couple of spins of the ol' seed wheel), to the more spicy level with genital ballsack exposure up to full genitals; although if you want to see full schmeat then I recommend a peen LoRA helper or use a checkpoint that knows how to generate 'em to get a first pass quality output.FLUX 2 Klein 9B version 1.2: Based largely out of the 4B dataset with some tuning - intent here is to be able to get the jockstrap right even with oblique angles as well. Like the 4B Klein version, this is very useful for image editing purposes ..Aside from full image generation, I often use it to edit existing images with edit mode with a simple prompt like: "1. put a {descriptive context} jockstrap on him 2. maintain everything else &lt;lora:this lora.safetensors&gt;"FLUX 2 Klein 4B version 1: My first Flux 2-Klein model! I decided to go with 4B for a couple of reasons.. one primarily how fast and yet high quality its outputs are, but more importantly the 4B model is under a Apache 2.0 (open source) model which allows commercial use of generated outputs.LTX-2 version 1: Since LTX-2 base doesn't know what a jockstrap is, I created this. It can be used for t2v or i2v.FLUX version 2: I thought since I created a new dataset for the ZIT version, consisting mostly of KREA/Z-Image/SRPO generated images using my previous Flux 1.2 version, I'd just toss in that ready dataset in the fast-Flux trainer, since now it only takes me a few minutes to complete, and see what came of it. The results were pretty good.Whereas the previous one was a lot of Internet scraped content that I cleaned/upscaled plus some SDXL/PDXL generations with all combined was a fairly large meticulously captioned dataset, this version has a much smaller dataset image in quantity, but much higher in quality, and with a very basic captioning strategy. The results are that this one is much better with prompt adherence and generally produces much higher quality outputs than v1.2 - however, since it was trained on a reduced dataset it doesn't have as much pose flexibility as the previous version. For example, rear/ass focused generations are hit or miss with mostly misses. I guess I didn't balance it out as well as I thought I did.That said, there are the benefits of higher quality and prompt adherence, especially with custom waistband logo specification and has some additional flexibility with ball cleavage easier (if prompted). It really didn't take long to make this version, so I'll eventually get around to making an updated one to address its ass gap (gape?) weakness. Overall not bad and in many respects better than previous version, so I thought I'd share.Z-Image Turbo version: My first foray into creating ZIT based LoRAs. The process was easy enough using AI Toolkit but I've noticed that ZIT is fairly sensitive with dataset quality and training step count. While I think the outputs are fairly good, although not as quality as I can get with Flux with a lot less fuss. That aside, I think it's decent enough to share, although it can get artifact-y with saturation and crispiness at times - if you encounter that try different samplers/schedulers.. I'm still tinkering with the new model training approach so if I figure out how to create a better LoRA result with it, I may end up creating an updated version.FLUX version 1.2: I created a FLUX LoRa for this clothing style concept because I wanted one that had a high degree of versatility in style, material, angle, model pose, and other factors. So this was trained against my previous SDXL/Pony large dataset, 800+ HD images, although cleaned and culled it a bit, and used a specific LLaVA prompt captioning parameters that focused on object + context approach. While I'm not happy with the size of the LoRA itself, it does the job I wanted it to do well so I thought I'd share it anyway.Used a cured dataset from previous Pony/SDXL versions, using a low and slow training approach. Dataset includes a fairly diverse set of jockstrap styles, poses while worn, angles, etc. with updated an Alpha 2 LLaVA captioning methodology.The captioning LLaVa AI was given specific prompt instructions to caption descriptions using a concept object-in-context strategy which focused on the jockstrap itself first, such as material, positioning, angles and other materially important attributes. Then other elements in the picture would be a secondary description priority. So, when using this LoRA a specificity such as waistband size, attributes about leg straps, pouch, etc., should work better than without.Why make yet another one when there are others available? The main goal with a low and slow + large dataset approach is to have a Flux LoRA that knows what a jockstrap is and not what you get typically get with a concept based Flux LoRA using a high and fast training approach: what jockstrap on guy looks like. The key goal/benefit with the former approach is that it really doesn't try to alter much, if at all, character elements. Don't get me wrong - I often use high and fast training approaches and Flux is really good learning that way - which is why character/people based Flux LoRAs are so easy to make.Power-bottom line: It plays well with character/person LoRAs and doesn't try to alter body/faces - just the jockstrap, ma'am, and nothing but the jockstrap.I'm using Replicate's LoRA trainer which gives quick access to a H100, at a very reasonable rate, but doesn't present many options as far as training parameters go - so if there's another site you suggest then let me know. My main gripe with my current approach is while very easy to create, the lack of parameter options means I can't optimize much so the LoRA size tends to be large. Since LoRA optimizers that work w/ SDXL/SD 1.5 LoRAs won't work with Flux based ones, here we are with the size.--Pony version: My latest version v5 was made to consolidate my previous version with a couple of my other conceptual LoRAs that involve jockstraps. The idea started with creating a LoRA with a few creative concepts without the difficulty of dealing with multiple ControlNets and/or a few very specific checkpoints. Plus many of the checkpoints generally know what a jockstrap is but often generate briefs without the use of a specific jockstrap based LoRA.Lastly, I also wanted a LoRA that knew what an actual jockstrap garment is, not just how it looks on a dude, instead of having to resort to a complicated workflow just to get one within an image by itself.So this one is trained not only improves on the variety and quality of generating guys wearing jockstraps in various angles , or used to inpaint, it can conceptually create concepts related to them.You can mix-and-match - it's not 100% but it shouldn't take many generations to get generally there.You can prompt with things like vantage point, color, actions or just the garment itself. The model gallery shows examples of the outputs and prompts used to create them to give you an idea.things like:[pov] [color] [smelling / sniffing] [on floor, face, holding, stretching] [peek] [no humans, display] etc.. you can combine them in various ways but the more complex the more patience will be required before you get what you want with a generation cycle.SDXL version 6: Is generally good for full image generation or inpainting. Prior versions are generally better for inpainting.Pony version: Suitable for full image generation or inpainting.<ul><li>RE: Pony + Inpainting: I found that Pony is generally more difficult to inpaint out of the box, but the right checkpoint, sampling, scheduler and LoRA it can be done without much of a mess. I've been using <a target="_blank" rel="ugc" href="https://civarchive.com/models/639170/virile-stallion">Virile Stallion</a> with fairly consistent results using either Euler A or DDIM with Karras using Fooocus and Krita. With A1111/ReActor I just use basically the defaults Inpaint Anything extension, or adjust sampling, schedule and checkpoint values using regular Img2img inpaint methods.</li></ul><hr />v1.2My first published LoRA - should be versatile enough for varied posesTrigger is jockstrap and you can specify varied colors - helps if you specify jockstrap front or back if you want to focus on a sideblack jockstrap, white jockstrap, purple jockstrap front, etc.jockstrap thong, jockstrap underwear, bulging jockstrap etc.

Jockstrap - varied colors, styles and concepts (Klein, LTX2, ZIT, FLUX, Pony & SDXL)

Adjusted rank/dim and other LoRA values to keep the general same output quality as original but down to about 35MB in size instead of 800+MB

Files with SHA256: b3229e5bafc1fb03ff944c4b38e10d7652e17812b765ee22a1902aff71b010e1