TLDR: IMPORTANT<ul><li>Use with <a target="_blank" rel="ugc" href="https://civarchive.com/models/2823254/kroma-krea-2-lora">https://civarchive.com/models/2823254/kroma-krea-2-lora</a> (both loras at strength 1)</li><li>online gen: remix this for quickstart: <a target="_blank" rel="ugc" href="https://civarchive.com/images/139792180">https://civarchive.com/images/139792180</a></li><li>If possible, gen at .5 megapixels and upscale (examples have workflows)</li><li>I'm really enjoying the dpmpp_2m_sde_heun_gpu sampler locally</li></ul>I went local, so I'm able to fiddle with more settings. This is a style lora trained against Kroma base v0.2. This Kroma version is still early in training and seems to struggle with higher resolutions. On the plus side, it already understands a lot of common booru and e621 tags.Lora training done with musubi at 736px resolution. Linear lr schedule, gradient accumulation steps 3, initial LR 0.00006, around 3.5 hours on a 5070 ti.

Files with SHA256: 42e45104e526532e41b9091920e568639c7a5d87c24057c4d54a4daa859cad12