<h1 id="!!-the-two-downloadable-lora-versions-behave-differently-!!">!! The two downloadable Lora versions behave differently !!</h1>The larger one (2.5GB) is the overall most accurate, but the smaller one (1.9GB) is more distilled and generalized. Meaning the smaller one is better used in combination with other loras and different styles as it carries less style/noise data and more the essential character. In direct comparision you can see that the smaller lora produces images with a clearly consistent style, while the larger lora varies wildly between prompts and seeds.<h1 id="the-lora-version-of-the-full-model-(v2.3-update)">The Lora Version Of The Full Model (V2.3 Update)</h1>V2.3 is rained on more than 6000 images, this lora achieves unparalleled generalization across everything that makes Disassembly Drones and Worker Drones destinct.I scraped literally every booru and image site and extracted and filtered every single one of the 300 000 frames of the ENTIRE show. I even finetuned a specialized CLIP model for the captions!Overall the distribution of the characters is wildly unbalanced, meaning: Very good at Uzi, V, J, Cyntessa, N, Doll and Lizzy; less good at Cyn and pretty bad at everything else. But with a specific enough prompt you can certainly bring every character from the Murder Drones universe on the canvas.I'm ngl this is a NSFW model first and foremost, so be warned!<h3 id="please-leave-a-like-and-a-comment-if-you-enjoy-my-models!">Please leave a like and a comment if you enjoy my models!</h3><hr /><h1 id="v2.3">v2.3</h1>A ground up retrain of the model on much more data, for longer and with multi caption.<ul><li>MUCH better character consistency and prompt following<ul><li>The model can generate MORE characters at once. Without mixups!</li><li>If mixups ocure, specify the hairstyle of the characters and that should give perfect accuracy.</li></ul></li><li>MORE freedom with artstyle<ul><li>Since the practically the entire show has been trained into this model you can switch two major artstyles!<ul><li>Illustration</li><li>3d</li></ul></li></ul></li><li>Best performance with a two stage workflow. You can get mine by downloading one of the example images.</li></ul><h2 id="v2.3:-comfyui-+-&quot;res_2s&quot;-+-sgm_uniform-+-cfg-4-+-30-steps">v2.3: ComfyUI + "res_2s" + sgm_uniform + CFG 4 + 30 steps</h2><h1 id="dataset-distribution">Dataset Distribution</h1>Number of images ≠ Quality of character (at least not always)<ul><li>UZIDOORMAN - 1960x</li><li>SD-N - 1897x</li><li>SD-V - 1590x</li><li>SD-J - 1140x</li><li>CYN - 546x</li><li>TESSA - 536x</li><li>DOLL - 271x</li><li>CYNTESSA - 242x</li><li>THAD - 172x</li><li>LIZZY - 164x</li><li>ALICE - 94x</li><li>NORI - 84x</li><li>KHAN - 75x</li><li>YEVA - 69x</li><li>SENTINEL - 64x</li><li>REBECCA - 22x</li><li>TESSA-ELIOT - 20x</li><li>DOLLS-DAD - 12x</li></ul><hr /><h1 id="v1.9">v1.9</h1><ul><li>Better character consistency and prompt following<ul><li>The model can generate many characters at once. Without mixups!</li><li>If mixups ocure, specify the hairstyle of the characters and that should give perfect accuracy.</li></ul></li><li>A little less freedom with the artstyle</li><li>Best performance with a two stage workflow. You can get mine by downloading one of the example images.</li></ul><h2 id="v1.9:-comfyui-+-&quot;er_sde&quot;-or-&quot;res_2s&quot;-+-sgm_uniform-+-cfg-4-+-30-steps">v1.9: ComfyUI + "er_sde" or "res_2s" + sgm_uniform + CFG 4 + 30 steps</h2><hr /><h1 id="v1">v1</h1><ul><li>Pretty good prompt following: Except it likes to mix up the topping and bottoming characters.<ul><li>Sometimes this can be fixed by just changing the order of them in the prompt</li></ul></li><li>Clothes arerelatively accurate, but you can change them to whatever you like</li><li>Best performance with a two stage workflow. You can get mine by downloading one of the example images.</li><li>It can do 3D, 2D and pretty much every other artstile Anima it self can do</li></ul><h2 id="v1:-comfyui-+-&quot;er_sde&quot;-+-sgm_uniform-+-cfg-4-+-30-steps">v1: ComfyUI + "er_sde" + sgm_uniform + CFG 4 + 30 steps</h2><hr /><h2 id="prompting-(for-v2.3):">Prompting (for V2.3):</h2><ul><li>Works best if you structure your prompt very well, so that every character tag can be cleanly associated with certain actions. Take my showcase images as examples!</li><li>If a character isn't generated correctly, type more describing traits: SD-V -&gt; SD-V with short white hair, robot<ul><li>BUT be aware that longer captions will ultimately degrade the models output! I noticed that this is the case with Anima in general.</li><li>Anima sometimes gives the drones human skin. Fix that by typing: robot<ul><li>(Tbh "robot" should almost always be included)</li></ul></li></ul></li><li>The Model sometimes forgets the tail. Luckily you can remind it by typing: tail</li><li>This time tag-based prompts are king, but natural language works too in some extend.<ul><li>Qwen3 is very flexible</li></ul></li><li>Sometimes the negative prompt can mess things up. Keep it minimal!</li><li>Always specify the amount of drones in an image!<ul><li>Like: 2 robots, robot, 1girl, 2girls, 1boy, etc</li></ul></li></ul>

AnimaDrones-V2.3-Lora

<ul><li>More data, More training, More filtering</li><li>Better at every character</li><li>Trained on the entire show (300k frames) filtered down to ~2k images.<ul><li>6k dataset images, ~12k captions (multi caption approach for character identity reinforcement and editability)</li></ul></li><li>Can be on model with good prompting</li><li>Short prompts work better than long prompts</li></ul>

Files with SHA256: 28b14cfc6f52a16c8baf65ec96e78f963ed007f8e54ecd6ac2ebf796b6b17392