So, let's get this out! It did not turn out as good as I expected, but why trash it 😅This LoRA is trained on UnNamedIXL V3 and do "something" to add back some unrealistic elements in the mix. It is highly biased toward cowboy shots and can mess up anatomy. It kinda do something with other realistic models.Helper keywords:<pre><code>cowboy shot, dynamic pose, dutch angle</code></pre>Recommended strength: no more than 0.8<h3 id="the-story-(tm)">The story (TM)</h3>So, my three flagships models handle this same prompt/seed/config like this:<img src="https://image.civitai.com/xG1nkqKTMzGDvpLrqFT7WA/c755523a-4f5f-4b19-8893-9d91cbeaa86f/width=525/c755523a-4f5f-4b19-8893-9d91cbeaa86f.jpeg" />The prompt was:<pre><code>IllusP0s, 1girl, solo, smug, realistic skin, green eyes, full-figured build, (neon pink hair, long flowing hair, twin braids), gold dancer outfit, jeweled bra, flowing skirt, golden bangles, exotic appearance, cowboy shot, eye catching, from above, dynamic pose
BREAK
desert, sand dunes, oasis, water pond, coconut tree, warm lights, clothe tents, volumetric lighting, subtle shadows, very aesthetic, cinematic, perfect hands</code></pre>As you can see, the base UnNamedIXL feels a bit flat and that is the case for most realistic models (some are exceptions and very good at more dynamic pose). So, the idea was to generate a bunch of pictures using Dynamic Prompts (aka Wildcards, built with ChatGPT help) with ShinV1 and then batch process those using Img2Img with UnNamedIXL V3 to train a LoRA.<img src="https://image.civitai.com/xG1nkqKTMzGDvpLrqFT7WA/c4ab8c56-3385-4e9b-893e-2e4bdbdd6085/width=525/c4ab8c56-3385-4e9b-893e-2e4bdbdd6085.jpeg" />All this was done using my DMD2 versions to do this quickly (a few hours instead of a few days for 200 pictures). But the training was done on the base UnNamedIXL V3. Here is the result of the process for the above picture:<ul><li>txt2img:</li></ul><img src="https://image.civitai.com/xG1nkqKTMzGDvpLrqFT7WA/9b1c71ed-e911-4402-bf19-a25c389a67f0/width=525/9b1c71ed-e911-4402-bf19-a25c389a67f0.jpeg" /><ul><li>Img2Img:</li></ul><img src="https://image.civitai.com/xG1nkqKTMzGDvpLrqFT7WA/bfe4138d-57e9-4c26-966e-79ed21aa628b/width=525/bfe4138d-57e9-4c26-966e-79ed21aa628b.jpeg" />Since the training data is heavily biased (mainly cowboy shot, it was my goal), i get an heavily biased LoRA ^^; (i have provided the training data with the LoRA if you want).I got 5 epochs like this:<img src="https://image.civitai.com/xG1nkqKTMzGDvpLrqFT7WA/76cfada5-444d-4ce3-ab4b-267afcc8e8a6/width=525/76cfada5-444d-4ce3-ab4b-267afcc8e8a6.jpeg" />Epoch 2 and 3 were the best (in my opinion) and mixing them is nice (both at 0.4) to avoid going too far from the base UnNamed IXL (see before) but still introduces some more flexibility.<img src="https://image.civitai.com/xG1nkqKTMzGDvpLrqFT7WA/26b27670-6069-41ef-a23b-b327a479139c/width=525/26b27670-6069-41ef-a23b-b327a479139c.jpeg" />So, i merged those like this (since i leverage "same to Strength" here, i lowered a bit the values):<img src="https://image.civitai.com/xG1nkqKTMzGDvpLrqFT7WA/30780aa0-6cad-4c5e-8210-d61a2771efd1/width=525/30780aa0-6cad-4c5e-8210-d61a2771efd1.jpeg" />To get this (at full strength):<img src="https://image.civitai.com/xG1nkqKTMzGDvpLrqFT7WA/8e5345be-2cc2-4f54-a409-6dae296aed6d/width=525/8e5345be-2cc2-4f54-a409-6dae296aed6d.jpeg" />Which was not perfect but i released anyway (with a nice "let's stick to 0.8" that nobody will listen to anyway 😉)Thanks for reading 🥰PS: merging Epochs from the same LoRA training is something i almost never see done or proposed, but i tend to do it often 😆I am pretty sure that block merging of LoRA could be the key to get overcooking out from same Epochs 😵Addendum: if someone wants epoch 4 or 5 since it landed the color of the bikini from the anime version, i'll release a mix of those too 😁And done, that's the "Spicier" version, i had to fix the hands a bit on top.<img src="https://image.civitai.com/xG1nkqKTMzGDvpLrqFT7WA/e50c68c1-bda5-4886-ad7d-e3037560799d/width=525/e50c68c1-bda5-4886-ad7d-e3037560799d.jpeg" />

Files with SHA256: 776da482aae3e030e3e8cfe9ba0a6152c0170cde28ef62b0beef460c53c223a7