<h2 id="flyingcumflux">flyingcumFlux</h2>This is my sixth 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.<hr />There are two versions of this LoRA. Both are recommended to be run at 1 to 1.2.<ul><li>01j is trained to 34 epochs. is an final stage LoRA that had success in generating a reliable erect penis ejaculating semen.</li><li>ema_99_flyingcumFlux01j20-32-34 is a Exponential Moving Average (ema) of epochs 20, 32 and 34. This merge generates a slightly different flying cumshot.</li></ul>I provide both LoRAs as I could not decide which to release.<hr />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:<ul><li>generate an erect penis ejaculating semen.</li><li>test whether captioning different fluid dynamics of semen ("semen physics") is trainable.</li><li>test whether hands holding penises is trainable</li></ul>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 an expanded set of images which showed a man ejaculating semen from different poses or photographic shots.<ul><li>The poses standing, sitting, kneeling, and reclining were equally represented. Reclining gens will change whether the image is in portrait or landscape orientation.</li><li>A subset of images contained hands free cumshots.</li><li>The tags were generated by joy-caption-beta-one.</li><li>Tags were manually edited.</li></ul>The primary activation tag "twinkcockFlux" was added to all images to maintain reverse compatibility with the other LoRA. A secondary tag (both the single word tag "cutflyingcumFlux" and the natural language phrase similar to "ejaculates a stream of semen from the tip" was added as a subconcept.Tags were manually edited to include descriptors of :Ejaculation length: "short length", "medium length", "long length"Ejaculation thickness: "thin", "thick" (not all captions contained this descriptor)Ejaculation direction: "upwards", "downwards", "arcing upwards", "arcing downwards" (not all captions contained this descriptor)Ejaculation quality: "white", "translucent white", "clear"Handedness: attempts were made to tag which hand was holding the penis. This was done from the perspective of the subject in the picture, "in his left hand", "in his right hand", "with both hands". A subset of images were tagged with "he is cumming handsfree".Sampler, Scheduling, Clip notes:<ul><li>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</li><li>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.</li><li></li><li>The LoRA was tested mainly on DEIS and DPM++2M (the sampler for civitai). DEIS is recommended.</li><li>Beta has been the most consistent sampler.</li><li>37-48 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 2.5-2.6.</li><li>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.</li><li>Using the original Clip-L model is recommended.</li></ul>Other notes:<ul><li>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.</li><li>The length or size of the penis, and the circumcision status of the penis was not consistently tagged in this LoRA.</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>Regularization images were used. Extensive testing on style flexibility beyond photorealistic has not yet been conducted.</li></ul>There were approximately 1794 images (including repeats and flips) used to generate this LoRA, at 1024x1024 resolution only.Special thanks to @markury 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.

flyingcumFlux

This is an EMA merge (0.99) of epochs 20-32-34.

Files with SHA256: ad89bd560c3b3b8bf6b28674925856ed02caef548f94b0716bb97b290c7f766d