⚔️ Nush from XTERNice Mommy is a NICE MOMMY <img src="https://image.civitai.com/xG1nkqKTMzGDvpLrqFT7WA/68c36e5b-8641-4ddc-b1de-2ef47b839b33/original=true/68c36e5b-8641-4ddc-b1de-2ef47b839b33.jpeg" /><hr />🧠<h3 id="trained-with-anima-standalone-trainer-(https:github.comgazingstars123anima-standalone-trainer)">TRAINED with <a target="_blank" rel="ugc" href="https://github.com/gazingstars123/Anima-Standalone-Trainer">Anima-Standalone-Trainer (https://github.com/gazingstars123/Anima-Standalone-Trainer</a>)</h3><ul><li>DIT model below</li><li><a target="_blank" rel="ugc" href="https://civarchive.com/models/2414435/animax-anima-finetune-model">https://civarchive.com/models/2414435/animax-anima-finetune-model</a></li><li>animaxAnimaFinetune_v05.safetensors</li></ul><hr /><h3 id="training-info">⚙️ Training Info</h3><ul><li>Trainer: Trained with <a target="_blank" rel="ugc" href="https://github.com/gazingstars123/Anima-Standalone-Trainer">Anima-Standalone-Trainer</a> on an RTX 3080 10GB.</li><li>Base DiT Model: <a target="_blank" rel="ugc" href="https://civarchive.com/models/2414435/animax-anima-finetune-model">animaxAnimaFinetune_v05.safetensors</a></li><li>TAGS - nush, xter,</li></ul><hr /><hr /><h3 id="dataset-and-training-details-(no-lies-edition)-for-v1.">📦 Dataset &amp; Training Details (No Lies Edition) for V1.</h3><h1 id="(v2epoch-23-checkpoint-is-much-improved-go-for-that-one-for-more-detailed-nush-look)">(V2[epoch 2/3 checkpoint] is much improved - go for that one for more detailed Nush look)</h1>Keeping it 100% transparent about the dataset quality so you know exactly what to expect:<ul><li>V1.0 details / for v2.0 check changelog notes</li><li>Dataset Size: 141 images.</li><li>Quality &amp; Variety: Varied! The dataset contains a mix of black and white sketches, manga panels, a few screenshots from hentai, doujinshi illustrations, and fan art.</li><li>Cleanup: Manually edited to remove watermarks, text tags, or background characters from the images.</li><li>Tagging: Pre-processed and auto-tagged using Kohya_ss (WD14 captioner with Threshold 0.5 / General threshold 0.35 / Character threshold 0.35 confidence settings), then manually reviewed to add missing details and delete conflicting tags.</li></ul><hr />❤️ Share Your Generations!

Anima - Nush [XTER]

🚀 OverviewVersion 2 is a complete ground-up overhaul focusing on high-fidelity facial consistency and micro-feature retention. By moving to an asymmetric dual-dataset structure and upgrading the network capacity, this version successfully fixes the recurring "vanishing mole" issue from previous versions, hard-baking it directly into the core character identity.🛠️ Key Improvements &amp; Fixes<ul><li>The Mole Fix (Hard-Baked Identity): Solved the inconsistent facial feature rendering from internal V1 / published V1 (internal V2). The signature mole is now permanently fused to the character trigger—no extra prompting required.</li><li>Facial Structure Smoothing: Eliminated the harsh, rigid "Chad-meme" jawlines found in early epochs/versions by optimizing total step duration.</li></ul>Enhanced Outfit Flexibility: High-resolution full-body images allow for robust clothing swaps without degrading facial accuracy.📊 Epoch Breakdown &amp; Warning LogWhen utilizing the generated safetensors, expect the following behavior per epoch:<ul><li>Epoch 1 (Under-baked / Avoid):<ul><li>❌ Lacks consistency.</li><li>❌ The mole does not always appear because the model hasn't logged enough steps to establish the facial texture.</li></ul></li></ul>⚠️🗿 Bug/Feature: Sudden, unprompted "Giga-Chad" masculine chin pop-ups. The model proudly defaults back to the ultra-defined base Anima structures before refining her actual feminine jawline. Not a bug—a feature for the bold.<ul><li>Epoch 2 &amp; Epoch 3[nush_improved_v2] (The Sweet Spot / Highly Recommended):<ul><li>NO MOLE TAG REQUIRED. It is a permanent, baked-in facial texture.</li><li>Smooth, accurate feminine jawlines and perfect facial structure.</li><li>Excellent outfit and background flexibility.</li></ul></li></ul>

Files with SHA256: d292c328ad86e70219eba6813f4eb51feb93a50831cd337c7a29c43cfd1852df