This works with CivitAI's on-site generator!!!

The Lora Version Of The Full Model

Trained on more than 4000 images, this lora achieves unparalleled generalization across everything that makes Disassembly Drones and Worker Drones destinct.

While I scraped literally every booru and image site, the training data has been processed down to only a few thousand images, with the best and most coherent asthetics and compositions. This theoretically means that every character in the entire show that has gotten any kind of fanart is in the dataset. Buuuut in reality not everything ever is posted to the big (or small) boorus.

Overall the distribution of the characters is wildly unbalanced, meaning: Very good at Uzi, V, J; Less good at Cyn and N 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!

Please leave a like and a comment if you enjoy my models!

v1

• Pretty good prompt following: Except it likes to mix up the topping and bottoming characters.

  • Sometimes this can be fixed by just changing the order of them in the prompt

• Clothes arerelatively accurate, but you can change them to whatever you like

• Best performance with a two stage workflow. You can get mine by downloading one of the example images.

• It can do 3D, 2D and pretty much every other artstile Anima it self can do

v1: ComfyUI + "er_sde" + sgm_uniform + CFG 4 + 30 steps

Prompting:

• Works best if you start the prompt with the character tag: SD-V, SD-J, UZIDOORMAN, CYN, etc

• If a character isn't generated correctly, type more describing traits: SD-V -> SD-V with short white hair, robot

  • SD-N often gets mixed with J for some reason

  • Anima sometimes gives the drones human skin. Fix that by typing: robot

    • (Tbh "robot" should almost always be included)

• The Lora likes to forget the tail. Luckily you can remind it by typing: tail

• This time tag-based prompts are king, but natural language works too in some extend.

  • Qwen3 is very flexible

• Sometimes the negative prompt can mess things up. Keep it minimal!