<h2 id="twinkcockflux-(alpha43000)-qxpvfqxbe">TwinkCockFlux (alpha43000)</h2>This is my third major publicly posted LoRA. This is a continuation of me learning how to use large language models (LLM) to generate an effective LoRA.This is the version that was trained to 43000 steps. At this step number, at times the LoRA will fail to generate a penis, or will be slightly blurred. Batch generation is recommended. Beyond this step number the LoRA became overtrained and began to at times lose prompt adherence, increased hallucinating, or render a penis to such a degree that it was no longer aesthetically appealing.This is a concept LoRA that produces a penis using the Flux_1.dev model. The goals of this LoRA were to:<ul><li>generate a circumcised penis</li><li>generate a circumcised penis interacting with clothing</li><li>test the feasibility of tagging the same image with categorically different tags</li></ul>I consider this an alpha because like TwinkCockXL, it works most of the time, but still is not perfect.This LoRA was trained on the basic set of images from TwinkCockXL but has some notable additions which are:<ul><li>More representation of Asian, Black and Latino young men. Addition of older men that are "post-twink". The majority of images still are white twinks. This LoRA appears more context dependent and less age sensitive than TwinkCockXL.<ul><li>Added comment: on further testing hair color and lighting seems to have an interaction with non-White generations. Hair colors other than black or brown may skew the generation away from Asian, Black or Latino characters. Extreme low light may cause the LoRA to confuse Black and South Asian generations. Adding "a dim white light above his head" may fix the problem.</li></ul></li><li>A different tagging strategy to overcome the 70 tag word limit of Stable Diffusion, where each image was tagged three different ways which were:<ul><li>Penis: Describing the penis (ethnicity, erection state, size, special interactions with clothing)<ul><li>{white,black,latino,asian}; {flaccid, semi-erect, erect}, {small-sized, average-sized, large-sized}</li></ul></li><li>Guided: Describing the main character of the photograph (ethnicity, body type, hair color and style, clothing)</li><li>Generalist: Describing the photograph in general (Description of photograph, angle of shot, the time of day the photograph was taken, the decade when the photograph was taken, what kind of camera was used to take the picture, the lighting and color tone of the photograph)</li></ul></li><li>The addition of "Flying Cum" as an experimental concept of the "magic moment".</li></ul>The primary activation tag "twinkcockFlux" was added to all images. A secondary tag (both the single word tag and the natural language phrase) was added to the following subconcepts:<ul><li>"cutbetweenpantsandshirtFlux", "his penis is between his pants and shirt" - a man wearing a shirt and bottom wear (pants, shorts, underwear), with a cut penis visible.</li><li>"cutcockthruflyFlux", "his dick is poking through his fly", "his penis is poking through his fly" - a man with a cut penis visible through his fly.</li><li>"cutcockthrulegholeFlux", "his penis is slipping out of his leghole", "his penis is poking out of his leghole" - a man with a cut penis visible through the leghole of his bottom wear, not consistent, may need to specify shorts.</li><li>"cutnopantsFlux", "he is not wearing pants" - a man with a cut penis visible while wearing a shirt but no pants.</li><li>"cutnoshirtFlux", ""he is not wearing a shirt" - a man with a cut penis visible while wearing bottom wear but no shirt.</li><li>"cutfullynudeFlux", "he is fully nude"- a man with a cut penis visible while not wearing any clothing.</li><li>"cutflyingcumFlux", "He is holding his erect penis and having an orgasm" also "he is holding his erect penis, he is having an orgasm, he is jerking off, he is cumming, he is masturbating" - a man with a cut penis at the "magic moment".</li></ul>The use of secondary tags is not necessary if randomness is desired. If a subconcept is desired, unlike twinkcockXL, Flux appears to be more prompt compliant after adding both the single word secondary tag and the natural language phrase. Secondary tags should not contradict each other for consistent generations.The cutflyingcumFlux data set was cross-tagged with other secondary concepts, which at times leads to generations having a "unexpected ejaculations" even if not explicitly specified. I am on the fence as to whether this is a flaw or a feature of the LoRA. Prompts with the subject sitting down, or holding his erect penis, will more likely generate this subconcept due to the source images for this subconcept. This secondary concept is experimental and is not consistent, recommend a strength of 1.2.Sampler, Scheduling, Clip notes:<ul><li>Initial testing was done using Comfy-UI and a custom workflow. Late in sample generation for this post it was discovered that ForgeUI generated more consistent generations than the Comfy-UI workflow, likely because of how the Comfy-UI workflow was programmed.</li><li>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).</li><li>For samplers, heun, euler, ipndm, deis generate similar results with heun and deis being the preferred sampler. dpm_adaptive in comfy-UI generates a different generation compared to the others from the same seed so often generating both or several generations (e.g. XY plotting) is recommended. Using dpm_adaptive was not successful with Forge UI.</li><li>Beta has been the most consistent sampler.</li><li>38-58 steps has been the most consistent, but depends on the sampler and the other settings.</li><li>Perturbed Attention Guidance for testing was set to 2.60-2.95.</li><li>Max shift was set to 1.15, base shift was set to 0.5.</li><li>Due to hardware specs, most of the testing of this LoRA has used the flux1-dev-Q5_K_S model.</li><li>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.</li></ul>Other notes:<ul><li>Unlike twinkcockXL, specific training on age ranges was not included in twinkcockFlux.</li><li>Similar to twinkcockXL, twinkcockFlux has "he is wearing a necklace", "he is wearing airpods" "there is sunlight across his body", "he is wearing a baseball cap" concepts at a high enough number to consistently generate.</li><li>Approximate era when the photograph was taken was added (e.g. "shot in the 1970s", 2020s). This had the unintended consequence of decreasing the image quality of a digital photograph to the approximate era (e.g., a low-resolution early 2000s digital camera with increased blur). If you encounter blur, consider explicitly specifying "shot in the 2020s" or "taken in the 2020s".</li><li>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.</li><li>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.</li><li>Images of backsides without a visible penis were not included in the dataset.</li><li>A few images of multiple men were included in the dataset but not at a high enough number to consistently generate high quality images.</li><li>Regularization images were used, but were a very small percentage. Extensive testing on style flexibility beyond photorealistic has not been conducted.</li></ul>There were approximately 14800 images (including repeats and flips) used to generate this LoRA, at 1024x1024 resolution only. Repeats were used to balance the number of each secondary concept so that they were generally equally represented.Special thanks to @markury, @spiritparticle and @wolffur666456 and the members of the Bulge Discord server <a target="_blank" rel="ugc" href="https://thebulge.xyz">https://thebulge.xyz</a> for their support, advice, and beta testing.

TwinkCockFlux_alpha

This LoRA was trained to 43000 steps, and is a balance between generation quality and prompt adherence.

Files with SHA256: 5ce0a9ce11d9a4dab8679464d3f3aa03c73a8dd35e9276c1c6e31c2563789e33