<h1 id="arturo-wolff-character-lora-for-anima">🐺 ARTURO WOLFF — Character LoRA for Anima</h1><blockquote>A character LoRA built to reproduce Arturo Wolff —me.</blockquote>This is my Arturo Wolff character LoRA, trained natively for Anima.Arturo is my personal character, online identity, and the central figure behind most of the artwork I make. Over time, his design has accumulated a lot of little visual decisions that are easy to describe individually but surprisingly difficult to reproduce as a coherent whole: the proportions, facial structure, scruffy silhouette, color separation, eyes, ears, tail, posture, expressions, and that permanently exhausted-looking rockstar energy.The purpose of this LoRA is to compress all of that into a reusable character representation.The goal is:Arturo should remain Arturo even when everything else changes.<hr /><h2 id="who-is-arturo">◆ WHO IS ARTURO?</h2>Arturo Wolff is a slim anthropomorphic wolf with a deliberately recognizable visual identity:<ul><li>Gray fur</li><li>Dark / black countershading</li><li>Black ears</li><li>Red eyes</li><li>Scruffy fur</li><li>Large pointed wolf ears</li><li>Digitigrade anatomy</li><li>Large, fluffy two-tone wolf tail</li><li>Lean proportions rather than an exaggerated muscular build</li><li>Naturally tired-looking eyes and facial structure</li><li>Usually carrying some degree of dry, sleepy, unimpressed, or rockstar-like expression</li></ul>That last part is particularly important.The tired appearance is part of the character design.During dataset captioning, some of those identity-defining facial characteristics were intentionally left implicit rather than constantly being tagged as separate concepts. The idea was to encourage the network to associate them with Arturo himself, instead of learning that Arturo is simply a generic wolf who optionally receives a <code>tired_eyes</code> modifier.However, depending on the checkpoint, these tags might be necessary:<blockquote>“tired eyes, bags under eyes, notched ear, scruffy fur, slim, anthro, furry”</blockquote><hr /><h1 id="trigger-identity">◆ TRIGGER / IDENTITY</h1><h3 id="primary-identity-token">Primary identity token</h3><code>awffnima</code>For most prompts, I recommend treating the character token as the anchor and then describing the scene normally.You do not necessarily need to exhaustively reconstruct Arturo from tags every single time.That defeats part of the purpose of a character LoRA.A basic prompt can be as simple as:<code>awffnima, solo, black jacket, standing in a convenience store at night, looking at viewer</code>Then add pose, clothing, expression, framing, environment, style, artist tags, etc. as needed.If a particular generation begins drifting, explicit anatomical or color descriptors can be added to reinforce the relevant feature.<hr /><h1 id="example-prompts">◆ EXAMPLE PROMPTS</h1><h3 id="simple-character-test">Simple Character Test</h3><code>awffnima, standing, full body, looking at viewer, relaxed posture, black shirt, jeans, simple interior</code>A useful test because there is very little for the prompt to hide behind.<hr /><h1 id="prompting-philosophy">◆ PROMPTING PHILOSOPHY</h1>I generally recommend describing what you want Arturo to be doing, rather than spending half of the prompt repeatedly explaining who Arturo is.For example:Instead of:<code>gray wolf, black ears, red eyes, gray fur, black fur, fluffy tail, two-tone tail, scruffy hair...</code>try:<code>awffnima, sitting sideways on a diner booth, one arm resting over the backrest, holding a coffee cup, looking out the window</code>Then reinforce individual physical traits only when necessary.A good identity LoRA should perform some of the descriptive work for you.<hr /><h1 id="lora-strength">◆ LoRA STRENGTH</h1>It was meant to work only at 1.0 but you might have to up the antee in some cases.<hr /><h1 id="base-model">◆ BASE MODEL</h1>Architecture: Anima Training Base: CircleStone Labs — Anima Base v1.0 Model Format: Anima LoRA Implementation: Diffusers Network Module: <code>networks.lora</code>This was trained against Anima Base, which is the correct member of the Anima family for LoRA training rather than training directly against one of the more opinionated aesthetic/distilled variants.The LoRA can be experimented with on compatible Anima models and finetunes, but behavior may vary depending on how heavily the target checkpoint alters the original model.<hr /><h1 id="training-data">◆ TRAINING DATA</h1>Training images: 61*50 Regularization images: 0 Dataset repeats: 10 Training resolution: 1024 × 1024<hr /><h1 id="network-configuration">◆ NETWORK CONFIGURATION</h1>LoRA Rank / Network Dimension: <code>32</code> Network Alpha: <code>32</code><hr /><h1 id="optimization">◆ OPTIMIZATION</h1>Optimizer: AdamW 8-bit Weight Decay: <code>0.0001</code>Learning Rate: <code>0.0001</code> Model / Transformer LR: <code>0.0001</code> Text Encoder LR: <code>0</code>LR Scheduler: Constant<hr /><h1 id="caption-regularization">◆ CAPTION REGULARIZATION</h1>Caption Dropout: <code>0.05</code> / 5%A small amount of caption dropout was used during training.Conceptually, caption dropout discourages the network from relying perfectly on every caption element being present on every iteration. For character training, this can help reduce brittle relationships between the character and incidental tags in the dataset.The objective is not just:<blockquote><code>specific token combination = specific training picture</code></blockquote>but a more generalized internal representation of the character.<hr /><h1 id="precision-and-memory-configuration">◆ PRECISION &amp; MEMORY CONFIGURATION</h1>Training Precision: BF16 Gradient Checkpointing: Enabled Latent Caching: Disabled Full FP16: Disabled Noise Offset: <code>0</code>The exported LoRA itself contains:896 BF16 tensorswith approximately:69.27 million stored tensor elementsand a file size of approximately:132.3 MiB<hr /><h1 id="training-environment">◆ TRAINING ENVIRONMENT</h1>The embedded training metadata reports:Training software: <code>ai-toolkit</code> ai-toolkit version: <code>0.12.14</code> Training controller: Civitai Spine Controller GPU architecture: Ada Reported VRAM: 24 GBThe file also contains the standard SafeTensors training metadata necessary for model identification and reproducibility.<hr /><h1 id="what-i-want-from-this-model">◆ WHAT I WANT FROM THIS MODEL</h1><ol><li>For me to create content of my fursona</li><li>For others to create fanart of my fursona</li><li>Validation</li></ol><hr /><h1 id="about-the-character">◆ ABOUT THE CHARACTER</h1>Arturo exists at the intersection of my artwork, music, online identity, and the larger collection of characters and projects I've built around him.Arturo is also generally portrayed as a womanizer, depressed, whimsy, and other stuff I can't share here.You are absolutely allowed to take him out of that context tho.Put the wolf somewhere stupid.<hr /><h1 id="final-notes">◆ FINAL NOTES</h1>The model is meant to be used, pushed around, mixed with styles, given weird outfits, placed into environments completely outside of the original dataset, and generally stress-tested.If you make something interesting with him, I genuinely want to see it.🐺Have fun.

Arturo Wolff | Anima

Version 1Training images: 61 Training resolution: 1024 × 1024

Files with SHA256: 3968166bdf178d27c7aff30a32d4112a935cbfa6adffb13d379d057cacbaddee