---
license: apache-2.0
base_model:
- circlestone-labs/Anima
tags:
- peft
- lora
- adapter
---

# Experimental Anima LLLite Regional Controlnet

Apply Anima ControlNet-LLLite parameters: (https://github.com/kohya-ss/ComfyUI-Anima-LLLite)
![image](https://cdn-uploads.huggingface.co/production/uploads/68794279700df7682fb2a81f/VbflhTy1UENO7CbJG5p_M.png)

Region Color Mask + Output image: (Use basic colors for different region)
![ComfyUI_temp_dpavo_00008_](https://cdn-uploads.huggingface.co/production/uploads/68794279700df7682fb2a81f/T5AKs4ah5Zd_xMc7w6t9p.png)

# Usage
Use a color mask image as the conditioning input.  
Any color can be used to define a region. There is no fixed palette or strict RGB requirement.  
The mask background should be white.  
Using simple solid colors (red, green, blue, yellow, etc.) for different regions is recommended for clarity.  
The model was trained using manually masked conditioning images and therefore expects clearly separated regions.  

## Prompting

Normal prompting works fine.

However, the model currently **cannot determine which prompt corresponds to which colored region automatically**. It only receives the region mask as additional conditioning and does not perform explicit prompt-to-region matching.

To guide concepts into specific regions, use spatial prompts such as:  

* "girl on the left, cat on the right"  
* "character in the foreground, city skyline in the background"  

Combining this model with attention masking methods such as Forge Couple or the Attention Couple node can provide stronger prompt-to-region associations.  


![image](https://cdn-uploads.huggingface.co/production/uploads/68794279700df7682fb2a81f/DmpLmrOn5ukV_ytTgV1a2.png)

# Training
Trained on 580 images, 2 repeats, batch size 4, for 4400 steps  
Conditioning images were manually masked for each image  
Captions were generated with the help of [wd-eva02-large-tagger v3](https://huggingface.co/SmilingWolf/wd-eva02-large-tagger-v3) and https://github.com/pythongosssss/ComfyUI-WD14-Tagger custom node  
Trained using https://github.com/kohya-ss/sd-scripts 

# Limitations 
Because most training images consisted of close-up character compositions, generations involving distant subjects may not strictly adhere to the provided mask boundaries.



# Support
If you'd like to support future training runs and new experiments, you can support me here: https://ko-fi.com/sensou