S12 V4.2 update! | Fine-tune LoRA | Extended LoRA

<p><strong>This is an Extended LoRA and will require LOCON extension.</strong></p><p>This version comfortably works with 3d-ish(by default, or with positive realistc/3d), 2.5d-ish(weak 3d and realistic negative, or none), and classic 2d anime(strong negative 3d and realistic, like 1.2 weight and above).</p><p>Additionally this dataset has own images tagged "blurry,"</p><p><br />V4: Smoothing tagging further. More advanced training techniques, if available, like Weight Decay clipping.</p><p><strong>-You are here-</strong></p><p>v4 0.6: Adequate tagging and knowledge separation. Advanced Adan training with separated tenc and u-net weight decay. Complete basic negative tagging.</p><p><em>(Not Released)v4 0.5: Dataset expansion to over 500 images. Initial tagging of additional dataset. Advanced Adan training. Initial negative tagging.</em></p><p><em>(Not Released)v3.3.4: Better color theme tag. More tags. Optimized Adan training. Awesome drawn texturing. Stable. Experiencing interesting style of upscale.</em></p><p><em>(Not Released)v3.3.3 Themes: initial addition of color theme tag. Expanded tagging. New Adan training. Unstable.</em><br />3.3.2: Adan optimizer training v2. Complete tagging of base dataset. Trying to localize style to "ohwx", while separating abstraction elements.<br />v3.3.2 is trained as a fine-tune, so it will affect generation without trigger words, but some concepts are isolated in them, overall, it tries to not change contents of image in a major way, but sometimes can.<br /><em>(Not Released)3.3.1: Experiment with Adan optimizer. Went bad.<br />(Not Released)3.2: Itterating on settings, trying out Lion optimizer.<br />(Nor Released)3.1(V3 sv1): advanced tagging, still incomplete. Separation of some knowledge to specific words, like "abstract" and "powerful"<br />(Not Released)V3: Decent usable state, tagging incomplete. Custom deep mix trained.<br />(Not Released)V2: Something actually usable, moving to extended. ACM trained.<br />(Not Released)V1: just basic lora, learning to do stuff. ACM trained.</em></p>