I am a <s>deviant</s> gentleman of culture. I haven't been able to reliable add fertilization x-rays to my images(all the LORAs are crap). So my thought process went like this. What if I make a LORA for inpainting and manufacture a clean dataset? Well this is the LORA for inpainting. Only remains to make a shitload of poses with different characters and inpainting on them to make a good dataset for the <a rel="ugc" href="https://civarchive.com/models/77264">concept LORA</a>(but that is for the future).<h2>First of all check the instructions in the version Tab!</h2>Edit: 20230526: finished the concept LORA, if you don't need it for inpainting use <a rel="ugc" href="https://civarchive.com/models/77264">that </a>instead.Edit (hopefully final) 20230524 : Finished a stylized heart shaped impregnation LOCON. Seems to have good repeatability and seems rather reliable. With that one V4.1 I consider this project done and will move back to character LORAs and to doing a generalist impregnation LORA(which will likely be in the 4.1 style). Sadly reliability increases with the version so the first ones might be iffy.The LORAS specialize as following:<ul><li>V1, V2 and V3(192 and 512) for a semi 3d cropped style.</li><li>V3(128) and V4(Locon) for a semi stylized ovum surrounded style.</li><li>V4.1(Locon) stylized heart in the nucleus style.</li></ul>Edit (Again) 20230523 : Once again cleaned and simplified the dataset, this time tried making it a LOCON to see if i could catch more details. I was able to make it more reliable as long as it is a spherical cow in space or a mostly round or square inpaint mask of the appropriate size.I have another Iteration in the works but this particular project has me a bit burnt out.BTW: Leave me a review and some feedback will you! This was a lot of work.Ignore the crossed out stuff Below. I just finished V3 in 3 sizes 128, 192 and 512.I recommend to take the following precautions:<ul><li>Use the LORA which more closely matches the available space</li><li>192 seems to be an evil resolution(even if it is a multiple of 64), it you are going to use it remember to set the inpaint area to only masked and set the resolution to 192x192(it won't change the size of your image but will only pay attention to that area)</li><li>Using a wrongly sized LORa works but if the Lora is for a smaller resolution the image will be smaller or maybe repeated. If the Lora is bigger, the in painted part will be extra large and cropped.</li><li>Use denoising strength = 1 Anything below that has given me subpar results.</li><li>I recommend to use latent noise or latent nothing. Other setting work but YMMV.</li><li>Setting the inPaint area to only masked helps with distortion.</li><li>The lora seems stable in all the range but it is a bit hard to know if it is overfit due to it being rather simplistic.</li><li>Either use 128 or 512 unless you really need something in the middle then make sure to follow the suggestion for 192.</li><li>Check the color images I added to see more or less how big the inpaint image will look. Red is 128, Blue is 192, green is 256 and Yellow is 512.</li></ul>I pretty much restarted from scratch removing from my dataset all the speech bubbles(as i was unable to control their alignment). Also rotated everything to reduce orientation bias.Also added the 128 size LORA as that size is a lot easier to inpaint, as it is sometimes hard to find a larger space.In the gallery i will add the originals and the inpainted versions of the examples.I'll add the dataset as it might be useful in the future, also the tag summary but i doubt that one will be useful as i pruned most of it as they were false positives.<s>I seem to get good results using: &lt;lora:fertilization:1&gt;, fertilization, ovum, (sperm_cell:1.2)</s><s>Use latent noise or latent nothing otherwise the existing stuff will get on the way. I recommend to use it at denoise 1, if you want to lower it to get transparency YMMV.</s><s>Also be mindful of the space you give it, if it too small it might omit something.</s><s>*Edit: if you use it in a full image it works like a weird background.</s><s>*Edit2: Working on a V2, will probably consolidate everything under Fertilization as depending on the inpaint shape you can get both ovum and sperm or just one of them which can get annoying specially for batch inpaints. My other gripe is the size of the stuff in the inpaint, If I am not convinced with V2, I will tweak the dataset to make the images take less space of the training area to see if I can get it to try to be more compact.</s>

Fertilization for Inpainting

This is actually V4 but then again 2 and 3 while not bad were not a great improvement. To train this I modified the dataset to be 192x192. This way I was able to regularize the size of the image in the in painting zone. On the other hand if the zone is too big you might get duplicates or weird stretching. If there's demand I might try to make a 256 version. V1 is of course 512.I might give this another try doing adding vertical flip as i saw some bias. I am not sure about the speech bubbles of some of the dataset as they often appear pointing the wrong way(as expected).I recommend to use:<ul><li>Try to make the impaint area close to 192x192 or if smaller try making it either round or square.</li><li>Lora Weight: 1</li><li>imprint area: only masked</li><li>Denoising strength: 1</li></ul>

Files with SHA256: ea1f0a659db997690f7a52ba0de583c4231252845228d737809c12756d0de15f