Attention!, LoCon Version need install Extension: <a target="_blank" rel="ugc" href="https://github.com/KohakuBlueleaf/a1111-sd-webui-locon">https://github.com/KohakuBlueleaf/a1111-sd-webui-locon</a><h2>Update 2023-03-23</h2>add LoRA-v1.1<ul><li>traing by 100 images</li><li>resolution = "704,1024"</li><li>noise_offset = 0.05</li><li>lion optimizer</li><li>network_dim / network_alpha = 128</li><li>checkpoint: <a target="_blank" rel="ugc" href="https://huggingface.co/andite/anything-v4.0">Anything v4.5</a> (DO NOT Forget VAE)</li></ul><h2>Update 2023-03-08</h2>add LoCon-v1.1, smaller and better, only 2.5MB<h2>Update 2023-03-07</h2>add LoCon Version, it is small and great, give it try<h2>Original</h2>better than expected, Lora has great potential.<ul><li>checkpoint: <a target="_blank" rel="ugc" href="https://civarchive.com/models/9942/abyssorangemix3-aom3">abyssorangemix3-aom3</a></li><li>no trigger words required</li><li>recommond weight: 1.0</li><li>you can use any prompts</li><li>recommond use hires.fix</li></ul>The weight can be adjusted as desired. Besides the texture and style, the composition has a high degree of repetition, perhaps due to the insufficient number of images?I will keep researching.用了50张图训练，输出模型72MB。结果比我预想的要好，看来Lora的潜力很高。权重应该是可以根据需要随意调整，除了风格和纹理，构图重复度比较高，也许是我图片太少了？我会保持探索。

[LoCon/LoRA] Torino Style for AOM3A1

100 images were used for trainingnetwork_dim = 8, network_alpha = 2conv_dim=2, conv_alpha=1output is 12MB.

Files with SHA256: acf870af0d60523c7a387b3cc641ba06355ba3c0cac3206a722fa27f9fc64796