<h1 id="the-best-judy-hopps-lora-ever!">The best Judy Hopps LoRA ever!</h1><p></p><p>- Why it's paid?</p><p>- I spent 30k buzz and a ton of time preparing datasets and training a second version. And believe me, it worth paying!</p><h2 id="v2">V2</h2><p>This LoRA was trained on <strong>180 frames of Judy Hopps from Zootopia 2 </strong>and crops of important details like ear bun and clothes.</p><p>This time, I tried using natural language captions and a network-rank of 128-64, instead of 64-64. These are the main differences in training.</p><p><strong>Character trigger:</strong></p><pre><code>jdhpps</code></pre><h2 id="prompting">Prompting</h2><p>Use natural language as usual! Just call her <strong>jdhpps, </strong>that's it.</p><p>Add this for better style:</p><pre><code>The image is a high-quality, stylized 3D animated render, resembling a modern, high-budget cinematic CGI feature film. The artistic is focusing entirely on appealing, expressive, and polished animated stylization. The rendering utilizes advanced CGI techniques, featuring soft global illumination, precise ambient occlusion, and crisp cinematic rim lighting. Material textures are meticulously crafted with sharp detailed fur and rich physical materials—with a clean, vibrant, and perfectly optimized animated aesthetic.</code></pre><h2 id="canonical-outfits">Canonical outfits</h2><p>Undercover uniform:</p><pre><code>jdndrcvr, She wears an open dark blue jacket over a light blue v-neck shirt, dark pants, and a belt with a metal rectangular buckle.</code></pre><p>Dress:</p><pre><code>jdflrldrss, ears in a bun on the top of her head, She wears a yellow dress with a single bare shoulder and a subtle floral pattern, a large yellow flower at the opposite shoulder, and a band of small yellow flowers around the base of her ear bun.</code></pre><p>Mommy Judy:</p><pre><code>jdmmclths, She wears a pink sleeveless shirt with the text "don't worry be HAPPY", a pink heart-shaped glasses, a waist bag with a carrot print, purple and pink striped pants, a pink bow around the base of one ear, and a wristwatch.</code></pre><h2 id="ears-teeth-tail">Ears, teeth, tail</h2><pre><code>ears pointing up/down
ears in a bun on the top of her head
rabbit teeth visible
rabbit tail visible</code></pre><h2 id="known-issues">Known Issues</h2><p>NONE!</p><p></p><p></p><p></p><p></p><h2 id="v1">V1</h2><p>This LoRA was trained on <strong>111 frames of Judy Hopps from Zootopia 2</strong>.</p><p>I used the same 111-image dataset and the same Danbooru-style captions that I previously used for my NoobAI LoRA. Because of this, the model responds particularly well to <strong>Danbooru tag sequences</strong> and to its trigger word:</p><p><strong>Trigger:</strong></p><pre><code>jdhpps rabbit girl</code></pre><p>At the same time, the base model has a very good understanding of natural English, so regular descriptive prompts work perfectly well too.</p><h2 id="prompting">Prompting</h2><p>In my experience, the most reliable approach is to combine both formats:</p><ol><li><p>Start with a short sequence of Danbooru-style tags containing the trigger word and the important character-specific features.</p></li><li><p>Continue with a natural-language description of the scene, composition, lighting, atmosphere, camera angle, and any additional details you want.</p></li></ol><p>I tested the LoRA with both very short prompts and much longer, highly descriptive prompts covering lighting, materials, fine details, and image quality. The model handles both approaches well, and prompt length alone does not seem to produce a dramatic difference in the final result.</p><p>For a more consistent <strong>high-budget 3D animated movie look</strong>, I recommend adding the following block to your prompts:</p><pre><code>The image is a high-quality, stylized 3D animated render, resembling a modern, high-budget cinematic CGI feature film. The artistic direction strictly avoids photorealism, live-action photography, or uncanny realistic animal anatomy, focusing entirely on appealing, expressive, and polished animated stylization. The rendering utilizes advanced CGI techniques, featuring soft global illumination, precise ambient occlusion, and crisp cinematic rim lighting. Material textures are meticulously crafted to balance detailed tactile qualities—such as soft, plush stylized fur and rich physical materials—with a clean, vibrant, and perfectly optimized animated aesthetic.</code></pre><p>The example images attached to this model use several different prompting styles, so you can check their generation data for reference. The common pattern is that I usually begin with a tag sequence containing the trigger word and any relevant character-control tags, then describe the rest of the image more freely.</p><h2 id="character-details-and-useful-tags">Character Details and Useful Tags</h2><p>The LoRA learned several details directly from the training set.</p><p>It responds well to Judy's canonical outfit, including:</p><ul><li><p>her jacket</p></li><li><p>the shirt underneath</p></li><li><p>belt</p></li><li><p>black pants</p></li></ul><p>It also understands tags controlling several characteristic features, including:</p><ul><li><p>ears up</p></li><li><p>ears down</p></li><li><p>rabbit teeth</p></li><li><p>rabbit tail</p></li></ul><p>You can use these depending on which details you want visible in the image.</p><h2 id="flexibility">Flexibility</h2><p>Overall, the model is quite flexible.</p><p>It can generate Judy from different camera angles, with different expressions, poses, backgrounds, lighting setups, environments, and scene concepts while keeping her identity very consistent.</p><p>The face in particular was learned extremely well, and I am very happy with how reliably the LoRA preserves Judy's facial appearance.</p><p>Some more unusual compositions or scenarios can still be stubborn. If a particular idea does not work immediately, try generating several variations or slightly changing the wording and structure of the prompt.</p><h2 id="known-issues">Known Issues</h2><p>There are several limitations in this version.</p><h3 id="canonical-jacket-details-leaking-into-other-outfits">Canonical jacket details leaking into other outfits</h3><p>Judy's standard outfit became strongly associated with the character during training. When generating completely different clothing, the model may sometimes produce abstract shoulder patterns resembling the decorative elements of her canonical jacket.</p><p>This can happen even when the requested outfit has nothing to do with the original costume.</p><p>Lowering LoRA strength fixes this issue in some cases, but I don't recomend putting it below 0.7.</p><h3 id="incorrect-paw-pads">Incorrect paw pads</h3><p>The model occasionally adds hairless, skin-colored pads to Judy's fingers or palms. These details should not be present and are an artifact of the current training.</p><h3 id="the-tail-appears-too-often">The tail appears too often</h3><p>The LoRA has a strong tendency to include Judy's rabbit tail, sometimes even when it was never requested.</p><p>In some cases it may appear from an anatomically impossible position—for example, during a strict front-facing view where the tail should obviously be hidden.</p><p>Because <strong>Krea 2 Turbo does not support negative prompts</strong>, there is currently no particularly reliable way to suppress these artifacts through prompting alone.</p><p>I will likely address them in a future version of the LoRA.</p><h2 id="alternate-outfit">Alternate Outfit</h2><p>I also tried to teach the model another canonical appearance from the movie: Judy's <strong>yellow floral dress</strong>, including the distinctive hairstyle where her ears are wrapped into a bun-like shape on top of her head.</p><p>Unfortunately, these elements were represented much less consistently in the dataset, and the LoRA did not learn them reliably enough.</p><p>You can experiment with them if you want, but I would not currently recommend this LoRA specifically for reproducing that outfit.</p><h2 id="nsfw">NSFW</h2><p>Base LoRA can generate Judy without clothes, but won't show nipples or genitals. For explicit NSFW images you should use additional LoRA, for example <strong>SNOFS Krea 2</strong>.</p><h2 id="final-notes">Final Notes</h2><p>Overall, I am very satisfied with this version of the model. I especially like the image quality produced by Krea 2 and how well the LoRA preserves Judy's identity and facial features while still allowing a fairly wide range of scenes and compositions.</p><p>Do not expect every concept to work perfectly on the first generation. More complicated ideas may require several attempts and some prompt adjustment.</p><p>If you have trouble describing a scene, language models such as <strong>ChatGPT or Gemini</strong> can also be useful for turning an idea into a more detailed generation prompt.</p><p>Feel free to experiment, post your generations, and leave feedback. I am especially interested in seeing what kinds of scenes, outfits, and compositions people manage to create with it.</p>

Files with SHA256: c5fa56cc14385a97da07ba5d4a8dd55541bc86915b9fdd20bf68a439beef1d84