Flux Klein 9B v2:

I made this version with the large image dataset from my Flux 1.D 4.2 version, upgraded with a Flux 2 9B + Ultra Detail image power-wash (upscaled resolution, additional refined details, skin texture, color grading, etc.) and re-captioned with an ablated Qwen 3-VL caption model from a previous Joy based model. Results are ... pretty, pretty good (when it gets it right). Judge for yourself in the gallery - all are one-shot (cherry picked), no i2i edits, additional LoRAs and with standard Flux 2K9B defaults with step counts, sampler and a basic base BFL GGUF model.

That said, its stability is not perfect so you will get body horror about half the time - so when that happens just regenerate with a different seed or if its minor enough you should be able to fix with an easy minor i2i corrective-pass. Even with its partial stability, depending on the seed #, to me the benefits outweigh the frustration of trying to get the composition through other ways with fighting prompts and/or layering one or more ControlPoint processes in a workflow. I may work on a better version but for now the GPU rental cost the bake it that long right now isn't worth the butt-squeze.

Flux 1.D v4.2:

I created this to be a fairly versatile action based LoRA concept that can play well with stacked character LoRAs and from a high degree of angles, vantage points and scenarios. So I used a dataset of approx 1,050 images from previous PDXL/SDXL datasets, batch captioned them and salad tossed them in the LoRA training hopper and in about 25 minutes this fairly good gem.

I prior to creating this one, my previous Flux LoRA attempt and another I found on here were okay but there was a lot of body horror situations going on. I ended up just going with a txt2img Pony -> img2img Flux workflow which was tedious which is why I came to creating this. This one has a pretty good success rate with a one-shot txt2mg generation so long as your prompting is well structured and you may have to play with its strength (usually 1.0 works well but dialing it back helps too - I wouldn't go lower than 0.6).

This version can create full male anatomy but tends to not create peens exactly right. If that turns out to be the case w/ the checkpoint or prompt you're using, then consider a peen-helper LoRA or a img2img 2nd pass over the problem spots.