v4.0 <span style="background-color:rgb(26, 27, 30);color:rgb(193, 194, 197);font-family:-apple-system, BlinkMacSystemFont, &quot;Segoe UI&quot;, Roboto, Helvetica, Arial, sans-serif, &quot;Apple Color Emoji&quot;, &quot;Segoe UI Emoji&quot;;font-size:16px">Trained with only img2img dataset (52 images) and double as long (20 epochs). Works better with more elaborate prompts. Still have problems with glitchy anatomy. v3.0 <span style="background-color:rgb(26, 27, 30);color:rgb(193, 194, 197);font-family:-apple-system, BlinkMacSystemFont, &quot;Segoe UI&quot;, Roboto, Helvetica, Arial, sans-serif, &quot;Apple Color Emoji&quot;, &quot;Segoe UI Emoji&quot;;font-size:16px">This one is actually just merge of two prewious (v1.5 from civit trainer and v2 with low dimentions). <span style="background-color:rgb(26, 27, 30);color:rgb(193, 194, 197);font-family:-apple-system, BlinkMacSystemFont, &quot;Segoe UI&quot;, Roboto, Helvetica, Arial, sans-serif, &quot;Apple Color Emoji&quot;, &quot;Segoe UI Emoji&quot;;font-size:16px">Results are better, though it appears that process of merging two tensors together even without any modifications (pure math, mimiking what happens inside of lora loader) still hinders quality. I suppose such is the price for compacting everything into one file instead of two separate. <span style="background-color:rgb(26, 27, 30);color:rgb(193, 194, 197);font-family:-apple-system, BlinkMacSystemFont, &quot;Segoe UI&quot;, Roboto, Helvetica, Arial, sans-serif, &quot;Apple Color Emoji&quot;, &quot;Segoe UI Emoji&quot;;font-size:16px">If you want better results, use specified loras separately. v2.0 <span style="background-color:rgb(26, 27, 30);color:rgb(193, 194, 197);font-family:-apple-system, BlinkMacSystemFont, &quot;Segoe UI&quot;, Roboto, Helvetica, Arial, sans-serif, &quot;Apple Color Emoji&quot;, &quot;Segoe UI Emoji&quot;;font-size:16px">This ona was trained on my machine with ai-toolkit trainer. <span style="background-color:rgb(26, 27, 30);color:rgb(193, 194, 197);font-family:-apple-system, BlinkMacSystemFont, &quot;Segoe UI&quot;, Roboto, Helvetica, Arial, sans-serif, &quot;Apple Color Emoji&quot;, &quot;Segoe UI Emoji&quot;;font-size:16px">Set dimention and alpha to 4, used same dataset and 500 steps. <span style="background-color:rgb(26, 27, 30);color:rgb(193, 194, 197);font-family:-apple-system, BlinkMacSystemFont, &quot;Segoe UI&quot;, Roboto, Helvetica, Arial, sans-serif, &quot;Apple Color Emoji&quot;, &quot;Segoe UI Emoji&quot;;font-size:16px">Better holds general pose, but, expectedly, fails with details. <span style="background-color:rgb(26, 27, 30);color:rgb(193, 194, 197);font-family:-apple-system, BlinkMacSystemFont, &quot;Segoe UI&quot;, Roboto, Helvetica, Arial, sans-serif, &quot;Apple Color Emoji&quot;, &quot;Segoe UI Emoji&quot;;font-size:16px">General Main idea of this lora is to generate pictures of fisting, when arm is almost all the way inside of another person (preferrably - elbow not visible and shoulder close to hole). Still laking, but already able to generate good images. And only similar one I managed to fine on this site is <a target="_blank" rel="ugc" href="https://civarchive.com/models/1722532/lost-my-keys-deep-fisting?modelVersionId=1949329&amp;sync-account=green">this one </a>from user @<a target="_blank" rel="ugc" href="https://civarchive.com/user/HaydenHoosk">HaydenHoosk </a>but it is specifically tailored for generating furries (not my style, deffinetly). LoRA is trained with semi-syntetic dataset. Halve is made by passing real images from "orange youtube" through depth+openpose controlnets and other halve was generated with my supervision. I am still new in this field, so I don't quite understand wich settings of trainer and qualities of dataset impact what, but I will be improwing this model in the future.If anyone has good explanations or observations, I will be wery glad to hear them in the comments.

Shoulder deep fisting

Same dataset, different settings. Will watch, if it changes anything.

shoulder-deep_2(img2img+syntetic).zip

shoulder-deep-10(kohya).safetensors

v1.5(kohya)

Files with SHA256: 8dd85a3fab490f2f865c44f84271b1e7b34c488841febf46a550ea88868cfa48