AnIllustrious Multi-ControlNet LoRA

Generate anime artwork using the AnIllustrious v5 checkpoint with full support for Civitai style LoRAs (trained for Illustrious/Pony, not base SDXL) and multi-ControlNet pose/depth/canny control.

Fork of fofr/sdxl-multi-controlnet-lora with the base checkpoint swapped to AnIllustrious v5 so your favorite anime style LoRAs actually apply correctly.

Features

Anime-optimized base: AnIllustrious v5 (Illustrious/NoobAI-family SDXL finetune)
Style LoRAs: Load up to 2 simultaneous LoRAs with individual scales (lora_urls, lora_scales)
3-slot ControlNet: openpose, canny, midas depth, + more (independently controllable)
Flexible modes: txt2img, img2img, inpainting, with ControlNet on any of them
Optional refiner: SDXL refiner pass for final polish
Advanced controls: scheduler choice, guidance scale, inference steps, seed control

Usage

Basic (txt2img)

prompt: "1girl, happy, school uniform, masterpiece"
negative_prompt: "lowres, bad anatomy"
width: 832
height: 1216
num_inference_steps: 30
guidance_scale: 7.5
seed: 42

With LoRA

prompt: "1girl, happy, school uniform, masterpiece"
lora_urls: "https://civitai.com/api/download/models/..." (Hugging Face or Civitai URL)
lora_scales: "0.7"

Multiple LoRAs (separated by |):

lora_urls: "https://hf.co/.../lora1.safetensors|https://hf.co/.../lora2.safetensors"
lora_scales: "0.8|0.6"

With ControlNet (pose guidance)

controlnet_1: "openpose"
controlnet_1_image: <image with pose skeleton>
controlnet_1_conditioning_scale: 0.75

Stack multiple ControlNets:

controlnet_1: "openpose"
controlnet_1_image: <pose image>
controlnet_2: "canny"
controlnet_2_image: <edge map>

Credits

Base checkpoint: AnIllustrious v5 by Jedas
HF mirror: John6666/anillustrious-v5vae-sdxl
Upstream Cog: fofr/cog-sdxl-multi-controlnet-lora

License

Cog code: MIT (see LICENSE.md)
AnIllustrious checkpoint: faipl-1.0-sd (Fair AI Public License)

Ethical Use

This model is intended for personal creative projects. Please respect: - Artist attribution if using style LoRAs from specific artists - Content policies: do not generate malicious, illegal, or NSFW content - Copyright: generated images should not infringe third-party IP

Tips

Anime LoRAs work much better on AnIllustrious than on base SDXL — that’s the whole point of this fork
Start with lora_scale: 0.6–0.8 and adjust to taste
ControlNet + LoRA stacking can be overwhelming; try one at a time first
Refiner adds ~20% latency but often improves final quality