<p>Only for NoobAI anime models!</p><p>Make clothes less tight around pregnant belly if possible while minimizing style change.<br />Can work with pregnant belly sliders.</p><h1 id="v4-n33o1pp39"><strong><span style="color:rgb(21, 170, 191)">v4</span></strong></h1><p>v4 is basically v3 retrained on better resolution (832x1216) with better hardware (3090) in better precision (bf16). this somehow produce better belly shape for loosely draping or translucent clothing, but still has the same problem for hanging clothes. Also this version lessen the influence the body proportion.</p><h1 id="v3-zlrezm3ds"><strong><span style="color:rgb(21, 170, 191)">v3</span></strong></h1><p>v3 is trained with much less dataset, mostly to remove image with folded fabric under the belly.<br />also text encoder isn't trained anymore, no actual effect because i don't introduce new trigger.</p><p>v3 soft is v3 hard without in8 layer, make it less influence composition of the image.</p><h1 id="v2-7testzm14"><strong><span style="color:rgb(21, 170, 191)">v2</span></strong></h1><p>Add more images from danbooru as dataset to force draping on clothes.<br />Change the base model for lora training to <a target="_blank" rel="ugc" href="https://civarchive.com/models/833294?modelVersionId=1116447">NoobAI Eps 1.1</a>.</p><h1 id="v1-cdzktthiq"><span style="color:rgb(21, 170, 191)">v1</span></h1><p>Trained with AI generated images from <a target="_blank" rel="ugc" href="https://civarchive.com/models/1078146?modelVersionId=1330186">SteinIllustriousV4</a>, then train this lora on itself.</p><p>using kohya_ss with settings from <a target="_blank" rel="ugc" href="https://github.com/ThereforeGames/blora_for_kohya">blora_for_kohya</a>.</p><p>The training focused on in8 (layout), out0 (content), out1 (style) in the UNet, and layer0 of clip_2 (text encoder).</p><p>I use slicer to use only content, layout, and text encoder components, which is reflected in the 'clte' naming.<br />using this will decrease the dataset style influencing the generated image.</p>

[NoobAI] Loose Pregnant Clothing

<p>Trained it with full resolution (832x1216)<br />now in bf16 datatype<br />now with better belly shapes</p>

LoosePregV4.safetensors

v3 Soft

v3 Hard

v1.0

<p>Only for NoobAI anime models!</p><p>Make clothes less tight around pregnant belly if possible while minimizing style change.<br>Can work with pregnant belly sliders.</p><h1 id="v4-n33o1pp39"><strong><span style="color:rgb(21, 170, 191)">v4</span></strong></h1><p>v4 is basically v3 retrained on better resolution (832x1216) with better hardware (3090) in better precision (bf16). this somehow produce better belly shape for loosely draping or translucent clothing, but still has the same problem for hanging clothes. Also this version lessen the influence the body proportion.</p><h1 id="v3-zlrezm3ds"><strong><span style="color:rgb(21, 170, 191)">v3</span></strong></h1><p>v3 is trained with much less dataset, mostly to remove image with folded fabric under the belly.<br>also text encoder isn't trained anymore, no actual effect because i don't introduce new trigger.</p><p>v3 soft is v3 hard without in8 layer, make it less influence composition of the image.</p><h1 id="v2-7testzm14"><strong><span style="color:rgb(21, 170, 191)">v2</span></strong></h1><p>Add more images from danbooru as dataset to force draping on clothes.<br>Change the base model for lora training to <a rel="ugc noopener noreferrer nofollow" href="https://civitai.com/models/833294?modelVersionId=1116447" target="_blank">NoobAI Eps 1.1</a>.</p><h1 id="v1-cdzktthiq"><span style="color:rgb(21, 170, 191)">v1</span></h1><p>Trained with AI generated images from <a rel="ugc noopener noreferrer nofollow" href="https://civitai.com/models/1078146?modelVersionId=1330186" target="_blank">SteinIllustriousV4</a>, then train this lora on itself.</p><p>using kohya_ss with settings from <a rel="ugc noopener noreferrer nofollow" href="https://github.com/ThereforeGames/blora_for_kohya" target="_blank">blora_for_kohya</a>.</p><p>The training focused on in8 (layout), out0 (content), out1 (style) in the UNet, and layer0 of clip_2 (text encoder).</p><p>I use slicer to use only content, layout, and text encoder components, which is reflected in the 'clte' naming.<br>using this will decrease the dataset style influencing the generated image.</p>

&lt;p&gt;Trained it with full resolution (832x1216)&lt;br /&gt;now in bf16 datatype&lt;br /&gt;now with better belly shapes&lt;/p&gt;

<p>Only for NoobAI anime models!</p><p>Make clothes less tight around pregnant belly if possible while minimizing style change.<br />Can work with pregnant belly sliders.</p><h1 id="v4-n33o1pp39"><strong><span style="color:rgb(21, 170, 191)">v4</span></strong></h1><p>v4 is basically v3 retrained on better resolution (832x1216) with better hardware (3090) in better precision (bf16). this somehow produce better belly shape for loosely draping or translucent clothing, but still has the same problem for hanging clothes. Also this version lessen the influence the body proportion.</p><h1 id="v3-zlrezm3ds"><strong><span style="color:rgb(21, 170, 191)">v3</span></strong></h1><p>v3 is trained with much less dataset, mostly to remove image with folded fabric under the belly.<br />also text encoder isn't trained anymore, no actual effect because i don't introduce new trigger.</p><p>v3 soft is v3 hard without in8 layer, make it less influence composition of the image.</p><h1 id="v2-7testzm14"><strong><span style="color:rgb(21, 170, 191)">v2</span></strong></h1><p>Add more images from danbooru as dataset to force draping on clothes.<br />Change the base model for lora training to <a target="_blank" rel="ugc" href="https://civitai.com/models/833294?modelVersionId=1116447">NoobAI Eps 1.1</a>.</p><h1 id="v1-cdzktthiq"><span style="color:rgb(21, 170, 191)">v1</span></h1><p>Trained with AI generated images from <a target="_blank" rel="ugc" href="https://civitai.com/models/1078146?modelVersionId=1330186">SteinIllustriousV4</a>, then train this lora on itself.</p><p>using kohya_ss with settings from <a target="_blank" rel="ugc" href="https://github.com/ThereforeGames/blora_for_kohya">blora_for_kohya</a>.</p><p>The training focused on in8 (layout), out0 (content), out1 (style) in the UNet, and layer0 of clip_2 (text encoder).</p><p>I use slicer to use only content, layout, and text encoder components, which is reflected in the 'clte' naming.<br />using this will decrease the dataset style influencing the generated image.</p>

Files with SHA256: 5630bb3f3b8f8116c069d8a63642e3a3503463580fa8b9dbcb692c273f12e8cb