ReverseCowboyFlux (48000)

This is my fourth major publicly posted LoRA. This is a continuation of me learning how to use large language models (LLM) to generate an effective LoRA.

Please read the full LoRA notes to understand what this LoRA can and cannot do.

This is the version that was trained to 48000 steps. At this step number, at times the LoRA will fail to generate a coherent image or will be slightly blurred. Batch generation and adding a small amount of another Flux Penis LoRA at (~0.2) or a style LoRA at (~0.4) may help sharpen the image.

This is a concept LoRA that produces two men engaged in anal sex in a "Reverse Cowboy" position using the Flux_1.dev model. The goals of this LoRA were to:

• Generate two men engaged in a specific gay sex position.

• Test the feasibility of tagging an image with JoyCaption with 250 words, as Flux natively supports tags of over 500 words.

This LoRA was trained on a new set of images, it is limited in representation and was trained mostly on Caucasian White males in their early to mid-twenties.

Images were tagged with JoyCaption Alpha 2, using the Descriptive caption type and 250 Caption length. Information about lighting, camera angle, camera settings, composition style, depth of field, artificial or natural lighting were included. The model was also asked NOT to use ambiguous language. The caption style follows the general format of:

1. Describe generally what kind of image this is.

2. Specify the two men in the scene.

   1. One man should be described as "the bottom", "the man on the left/right (choose one)", "the man in the background"

   2. The other man should be described as "the man on top", "the man on the left/right (choose the other side)", "the man in the foreground"

3. In this description, please use "anal sex", "anally penetrating", "inserted into anus" "reverse cowboy position" as narrative text descriptions.

4. The tag "reversecowboyFlux" was included in all captions and can be used as an activation tag.

Sampler, Scheduling, Clip notes:

• Initial testing was done using ForgeUI.

• Recommended guidance ranges between 2.8 and 3.5. When generating the image samples for this post with Forge UI, 2.8 will generate more realistic lighting and contrast but will favor a blurry penis, 3.5 will generate sharper images overall but will become contrasty similar to high dynamic range images (HDRI).

• For samplers, DEIS and DPM++2M generate similar results.

• Beta has been the most consistent sampler.

• 38-58 steps has been the most consistent, but depends on the sampler and the other settings.

• Perturbed Attention Guidance for testing was set to 2.60.

• Max shift was set to 1.15, base shift was set to 0.5.

• Due to hardware specs, most of the testing of this LoRA has used the flux1-dev-Q5_K_S model.

Using the original Clip-L model is recommended, using other fine tuned clip models resulted in a noticeable drop in quality, increased hallucinations, and loss of prompt adherence.

Other notes:

• Like twinkcockFlux, specific training on age ranges was not included.

• Due to limitations on what was available when the training started, images were not masked for either faces or backgrounds. This LoRA will impact face generation and interact with other LoRAs.

• Women were not included in the training data, I do not plan on including women in the future as I do not have source images.

• No other sexual position was trained in this LoRA.

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

There were 922 images (including flips) used to generate this LoRA, at 128, 256, 512, 768 and 1024 resolutions only. Adafactor was used for adaptive learning rates.

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