cockunderFlux

cockunderFlux (ClearVPLdown)

This is my fifth major publicly posted LoRA. This is a continuation of me learning how to generate a LoRA, this time with cloud based services and kohya-ss. The data used to train this lora was also a marked departure from my other LoRAs.

Note: This LoRA is fairly temperamental and idiosyncratic as a dick print is essentially a shadow on clothing, and was one of the more difficult concepts to train. Factors that include successful generation include the lighting, background, and the type of clothing over the penis.

As always, batch generation is recommended.

---

There are two versions of this LoRA. Both are recommended to be run at 1.1 or 1.2.

• Ep-50 is trained to 50 epochs. is the final stage LoRA that had success in generating a robust dick print.

• ema_99_cockunderFlux_clearVPLdown_block4_micro_V01f-28_49_50 is a Expoential Moving Average (ema) of epochs 28, 49 and 50. This merge generates a slightly different dick print.

I provide both LoRAs as I could not decide which to release.

---

Summary for both LoRAs:

This is a concept LoRA that produces a penis using the Flux_1.dev model. The goals of this LoRA were to:

• generate a visible penis line through clothing with the flaccid penis in the downward position.

I consider this an alpha because like TwinkCockXL and with TwinCockFlux, it works most of the time, but still is not perfect.

This LoRA was trained on a highly curated set of images which showed a clear visible penis line.

• One focus of this lora was to represent different shot lengths, including cowboy shot (head to knees), medium full shot (head to thighs), headless torso shots (torso and crotch), headless shots including legs, and full body shots.

• The poses standing, sitting, crouching, and lounging were equally represented.

• The tags were generated by a mix of cogvlm-chat-hf and florence-2-large.

• Tags were manually edited

The primary activation tag "cockunderFlux" was added to all images. A secondary tag (both the single word tag "clearVPLdown" and the natural language phrase "showing a visible penis line in the downward position" was added as a subconcept.

Tags were manually edited to include "flaccid penis outline is visible through his clothing hanging to his {left|right} with a {straight on|partial left side|partial right side|left side|right side} view". Initial Testing has not indicated if this tagging was ultimately successful.

Initial testing has shown that the phrase "the lighting is dramatic" will more likely generate a clear VPL. As well, a closer look at the captions has suggested that "sheer thong", "underwear", "compression shorts", "sweat shorts", and "shorts" will more reliably generate a dick print.

Due to limitations on the training set images, directly specifying whether there are or are not tattoos may be needed.

Sampler, Scheduling, Clip notes:

• Initial testing was done on the Stability Matrix implementation of Forge UI. Testing was done using the flux1-dev-q5_k_m.gguf quantized model, and the flux1-krea-q5_k_m.gguf model

• Recommended guidance/Distilled CFG Scale for flux1-dev is between 2.8 and 3.5. Recommended guidance/Distilled CFG Scale for flux1-krea is 4.5.

• The LoRA was tested mainly on DEIS and DPM++2M (the sampler for civitai). DEIS is recommended.

• Beta has been the most consistent sampler.

• 37-48 steps has been the most consistent, but depends on the sampler and the other settings.

• Perturbed Attention Guidance for testing was set to 2.5-2.6.

• Due to hardware specs, most of the testing of this LoRA has used the flux1-dev-Q5_K_M and flux1-krea_Q5_K_M.

• Using the original Clip-L model is recommended.

Other notes:

• Unlike twinkcockFlux, due to limits on the training set, the LoRA was mostly trained on white Caucasian twinks. Ethnic and age representation will likely rely on the main Flux model. When this happens the generation is more likely to blur.

• The quality of the visible penis line will depend on the clothing over the penis. A "sheer thong" will result in the most clear print.

• Women were not included in the training data, and I have no clear idea what will happen if a woman is specified, One likely result may be a penis being added to any figure in the generation. I do not plan on including women in the future as I do not have source images.

• Regularization images were used. Extensive testing on style flexibility beyond photorealistic has not yet been conducted.

There were approximately 1440 images (including repeats and flips) used to generate this LoRA, at 1024x1024 resolution only.

Special thanks to @markury and @wolffur666456 and the members of the Bulge Discord server https://thebulge.xyz for their support, advice, and beta testing.