Geraldr.AI model

geraldraimodel.safetensors

v1.0

<h1 id="important-to-read:-r84a8s5h3">Important to read:</h1>(the workflow I usually use is on the last image of "ErectCir" if anyone wants to try it)<h2 id="erectcirv1:-a8c0tr8gb">ErectCirV1:</h2>Is recommended to use "erect circumcised penis" on your prompt.Be aware that this model can cause some deformations and has a tendency to make tanned skin, the tanned skin can be avoid by mentioning ethnicity, nacionalities, skin color, etc.(I decided to include this LoRA as a version of DynamicPenis since it was trained with that model as the base)<h2 id="dynamicpenisv2:-755wnf67h">DynamicPenisV2:</h2>Is important to use this sentence to avoid major issues on your outputs "very detailed flaccid/erected circumcised/uncircumcised penis" since this is the sentece used on the dataset.<hr /><h1 id="updates-and-recommendations:-58awq03dj">Updates and recommendations:</h1>I've been trying to improve the Dynamic penis model, but at this point I think is clear the wall we're hitting with Flux, although is very capable and versatile, when it comes to genitalia is a headache, but the main reason of struggle with training were the weird lines on LoRAs trained on Kohya_ss, I've tried many methods but finally found a solution, so for anyone struggling with that too, I'm sharing it:First of all, deactivate the option "Apply T5 mask attention", and using this model to train <a target="_blank" rel="ugc" href="https://civarchive.com/models/941929?modelVersionId=1319871">HERE</a> (just to clarify for this LoRA I still recommend the base model of Flux 1 Dev)Those lines commonly appear at higher resolutions, so as you can see in the examples, those were made on 1024x1440, and Flux seems to work better with 1440, and now with this model you can do higher upscales without those annoying lines too, I might recommend this LoRA too, since it helps already trained LoRAs, <a target="_blank" rel="ugc" href="https://civarchive.com/models/1049158?modelVersionId=1177215">HERE</a><h2 id="and-i-want-to-thank-everyone-for-the-support-less3-even-tho-i'm-not-a-pro-and-just-do-this-as-a-hobby-makes-me-happy-to-see-your-results-with-the-models-and-the-feedback.-wfjg7r3y4">And I want to thank everyone for the support &lt;3, even tho I'm not a pro and just do this as a hobby, makes me happy to see your results with the models and the feedback.</h2>Lastly, as I always say, this model is not perfect, it may give you some deformation sometimes, is not super versatile unfortunately but I'll continue work on that, specially once a model is done, I can see on paper what it lacks, so this is a continue discovering, test, try, fail, and maybe half success.

Dynamic Penis Flux

DynamicPenisV2_Flux.safetensors

v2.0

ErectCir

<h1 id="important-to-read:-r84a8s5h3">Important to read:</h1>(the workflow I usually use is on the last image of "ErectCir" if anyone wants to try it)<h2 id="erectcirv1:-a8c0tr8gb">ErectCirV1:</h2>Is recommended to use "erect circumcised penis" on your prompt.Be aware that this model can cause some deformations and has a tendency to make tanned skin, the tanned skin can be avoid by mentioning ethnicity, nacionalities, skin color, etc.(I decided to include this LoRA as a version of DynamicPenis since it was trained with that model as the base)<h2 id="dynamicpenisv2:-755wnf67h">DynamicPenisV2:</h2>Is important to use this sentence to avoid major issues on your outputs "very detailed flaccid/erected circumcised/uncircumcised penis" since this is the sentece used on the dataset.<hr /><h1 id="updates-and-recommendations:-58awq03dj">Updates and recommendations:</h1>I've been trying to improve the Dynamic penis model, but at this point I think is clear the wall we're hitting with Flux, although is very capable and versatile, when it comes to genitalia is a headache, but the main reason of struggle with training were the weird lines on LoRAs trained on Kohya_ss, I've tried many methods but finally found a solution, so for anyone struggling with that too, I'm sharing it:First of all, deactivate the option "Apply T5 mask attention", and using this model to train <a target="_blank" rel="ugc" href="https://civitai.com/models/941929?modelVersionId=1319871">HERE</a> (just to clarify for this LoRA I still recommend the base model of Flux 1 Dev)Those lines commonly appear at higher resolutions, so as you can see in the examples, those were made on 1024x1440, and Flux seems to work better with 1440, and now with this model you can do higher upscales without those annoying lines too, I might recommend this LoRA too, since it helps already trained LoRAs, <a target="_blank" rel="ugc" href="https://civitai.com/models/1049158?modelVersionId=1177215">HERE</a><h2 id="and-i-want-to-thank-everyone-for-the-support-less3-even-tho-i'm-not-a-pro-and-just-do-this-as-a-hobby-makes-me-happy-to-see-your-results-with-the-models-and-the-feedback.-wfjg7r3y4">And I want to thank everyone for the support &lt;3, even tho I'm not a pro and just do this as a hobby, makes me happy to see your results with the models and the feedback.</h2>Lastly, as I always say, this model is not perfect, it may give you some deformation sometimes, is not super versatile unfortunately but I'll continue work on that, specially once a model is done, I can see on paper what it lacks, so this is a continue discovering, test, try, fail, and maybe half success.

Files with SHA256: 7b7e8f279af83fa180d17edb0ced4df33fc75fd3353dadbb4deadaeb0dbd910a