An experiment on traning the w5 layer block (from blora paper), but with loha instead of lora.<hr />!!!<ol><li>This loha is trained with fp8 unet without training text encoder. So, probably it will be bad.</li><li>You can download the image and drag n drop to your comfyui for more detail on prompt, sampler, etc;</li><li>I'm using either <a target="_blank" rel="ugc" href="https://github.com/Kim2091/Kim2091-Models/releases/tag/2x-AnimeSharpV2_Set">2x-AnimeSharpV2_ESRGAN_Soft</a> or lanczos for upscaling (hires fix).</li></ol>!!!<hr />notes:<ul><li>v1<ul><li>Lora str = -1 - 1;</li><li>Trigger word = tpppt;</li><li>Style : hand drawn black and white pencil;</li><li>Recommended checkpoint : hassakuXLIllustrious_v21fix;</li><li>Recommended sampler : dpmpp 2m karras with hires fix.</li></ul></li><li>v2<ul><li>Lora str = 0 - 1.25 (play with it, depends on models and prompts. go above 1 if necessary);</li><li>Trigger word = caststatsimple;</li><li>Style : flatish color and chibi/simple artstyle?;</li><li>Recommended checkpoint : hassakuXLIllustrious_v21fix;</li><li>Recommended sampler : euler a (normal) with hires fix.</li></ul></li></ul><hr />logs:<ul><li>v1<ul><li>Trained on illustriousXL_v01;</li><li>Data : 5 img by the artist <a target="_blank" rel="ugc" href="https://danbooru.donmai.us/posts?tags=tp_p_pt+&amp;z=5">tp p pt</a>;</li><li>// Initial concept, testing and training.</li></ul></li><li>v2<ul><li>Trained on illustriousXL_v01;</li><li>Data : 40 img by the artist <a target="_blank" rel="ugc" href="https://danbooru.donmai.us/posts?tags=channel_%28caststation%29+&amp;z=5">caststation</a>;</li><li>// testing with bigger datasets.</li></ul></li></ul><hr />sry if my english is bad, not a native speaker.

Files with SHA256: e0c034e5670bf4f4c6eeaddccfe44832cfbafa19f0b031f59cba6e12f5659656