<a target="_blank" rel="ugc" href="https://www.buymeacoffee.com/pettankoutei"><img src="https://image.civitai.com/xG1nkqKTMzGDvpLrqFT7WA/e4a77da5-bdb6-47ef-adfb-2372e692cc6d/width=525/e4a77da5-bdb6-47ef-adfb-2372e692cc6d.jpeg" /></a>If you like my content consider supporting me! 😊1 Cookie = Early access to future releases! 🍪❤️Locon trained on the style of Yamamoto Souichirou for Pony Diffusion v6 XL based models.<h3 id="heading-112">Pony versions:</h3>All sample images uses a weight of 0.8 and no trigger word.Since it is prominent in this specific style, the tag 'forehead' makes the style closer to the original. Also, the tag 'grin' is highly recommended 😊.In my opinion the end result looks very close to the artist's illustrations, but I think sometimes it might negatively affect body composition or hand/feet anatomy. For that reason I made 'special mix' version: a couple of community Loras of similar style at very low weights merged to the original trained model. The end result is very slightly different, but I think I managed to improve the aspects mentioned while retaining 95% of the original style. Personally I'm statisfied with both results, so I'll be releasing both versions.To sum it up:<ul><li>V1.0: closer to the original illustrations style. Might have some hand/anatomy problems. Pick this one if you want to use it alone or mix with other styles.</li><li>V1.0 Mix: still very very close to the original style, but in my opinion a bit more stable.</li></ul><h3 id="heading-113">SD1.5 version:</h3>This version uses a weight of 1.0 and no trigger word.Again, since it is prominent in this specific style, the tag 'forehead' makes the style closer to the original.The SD1.5 version learned very well the style in terms of characters proportions and composition, but it doesn't emulate the look of Yamamoto Sensei's lines or coloring style. Instead, it seems to adapt to the style of the model used, so I guess it ended up quite flexible in that aspect.To try achieving the typical look of his illustrations I used <code>(traditional media, sketch:1.25)</code> when testing with the AnyLora checkpoint (since it is closer to NAI and wouldn't influence much the model). Weighting <code>sketch</code> too high makes the background too simple, so either reduce weight or try to counter it with prompts like <code>detailed background</code>.I'm still experimenting with training parameters for Pony and SDXL, so if you have any tips or more advanced knowledge for better training with Pony let me know through Chat/DM.<h2 id="heading-1488">Disclamer</h2>All the models that I have developed and published on this platform are for entertainment purposes only. They are derived from the base Stable Diffusion model, which means it also includes its inherent issues, limitations and unfiltered capabilities. With that in mind, I am in no way responsible for any content that the user creates with malicious intent while using any of these models. The user assumes complete responsibility for the misuse of these tools and any unethical content created with them. This disclaimer is subject to the laws of the United States of America and the state of California. I reserve the right to remove or report any content that violates the terms of service, ethical standards, or applicable laws.<h2 id="heading-1489">License conditions</h2><a target="_blank" rel="ugc" href="https://civarchive.com/models/347550">Yamamoto Souichirou Style for Pony Diffusion v6 XL and SD1.5 </a>© 2024 by <a target="_blank" rel="ugc" href="https://civarchive.com/user/pettankoutei">Pettan Koutei </a>is licensed under <a target="_blank" rel="ugc" href="http://creativecommons.org/licenses/by-nc-sa/4.0/?ref=chooser-v1">Attribution-NonCommercial-ShareAlike 4.0 International.</a>

Yamamoto Souichirou Style

🔍 Version info:Initial release. Trained 200 images for 2700 steps.📋 Training Parameters:<ul><li>Base Model: <a target="_blank" rel="ugc" href="https://civitai.com/models/257749?modelVersionId=290640">Pony Diffusion V6 XL</a></li><li>Training Resolution: 1024</li><li>Training Batch Size: 1</li><li>Type: LoCon</li><li>DIM / Alpha: 8 / 4</li><li>Conv Dim / Alpha: 24 / 6</li><li>Train On: Unet + TE</li><li>Optimizer: Prodigy</li><li>Min SNR Gamma: 8</li><li>Bucket Resolution: 512 - 2048</li><li>Bucket Steps: 64</li><li>Noise Offset: Normal (0.035)</li><li>Image Count: 200</li><li>Steps: 2700</li></ul>

Files with SHA256: 1d866795c34c710e15f81e8eaa78cdef3d4aa7a6d8d2cec2c7fb32f890c62578