SubMaroon/ControlNet-manga-recolor

ControlNet for Manga Colorization

Model Name: SubMaroon/ControlNet-manga-recolor
Base model: John6666/nsfw-anime-xl-v1-sdxl
Task: Conditional image generation — Colorization
Conditioning: Grayscale manga panel (lineart or filled)
Trained with: Hugging Face diffusers ControlNet training pipeline

---

Description

This is a custom-trained ControlNet model designed to perform automatic colorization of grayscale anime styled images.

The model takes in a black-and-white anime styled pictures (converted to RGB) as conditioning input and generates a colorized version using Stable Diffusion.

It is trained to act as a ControlNet module and requires a compatible SDXL base model — such as nsfw-anime-xl-v1-sdxl or other anime/manga-focused SDXL models.

---

Training details

• Base model: John6666/nsfw-anime-xl-v1-sdxl
• Dataset: Custom dataset of ~6,000 image pairs from Danbooru-based manga scans, manually cleaned and resized to 768x768
• Inputs:
  • conditioning_image: black-and-white manga scan (RGB)
  • text prompt: optional (e.g. "1girl, blue_eyes, blue_hair etc.")
• Loss: MSE with FP16, trained on 1×RTX3090, 4 epochs
• Resolution: 768x768
• Scheduler: default diffusers setup
• Optimizer: LR: 1.4e-4

---

Usage (Diffusers)

from diffusers import StableDiffusionControlNetPipeline, ControlNetModel
from diffusers.utils import load_image
import torch

# Load ControlNet
controlnet = ControlNetModel.from_pretrained("SubMaroon/ControlNet-manga-recolor", torch_dtype=torch.float16)

# Load base pipeline
pipe = StableDiffusionControlNetPipeline.from_pretrained(
    "John6666/nsfw-anime-xl-v1-sdxl", controlnet=controlnet, torch_dtype=torch.float16
)

pipe.to("cuda")

# Load grayscale manga panel
conditioning_image = load_image("bw_manga_panel.png").convert("RGB")

# Generate
image = pipe("manga colorization", image=conditioning_image, num_inference_steps=30).images[0]
image.save("colorized.png")

---

Usage in ComfyUI / WebUI

• Place diffusion_pytorch_model.safetensors into your ComfyUI/models/controlnet/ folder
• Make sure to also include the config.json
• Select this ControlNet in your workflow
• Use grayscale images as conditioning inputs

---

Alternative training run (SDXL version)

This version was trained using the SDXL-compatible ControlNet pipeline with the following CLI command:

accelerate launch train_controlnet.py \
  --pretrained_model_name_or_path="John6666/nsfw-anime-xl-v1-sdxl" \
  --dataset_name="SubMaroon/danbooru-colored" \
  --image_column="image" \
  --conditioning_image_column="conditioning_image" \
  --caption_column="prompt" \
  --output_dir="./controlnet-colorization" \
  --resolution=768 \
  --train_batch_size=4 \
  --gradient_accumulation_steps=4 \
  --learning_rate=1.4e-4 \
  --num_train_epochs=12 \
  --mixed_precision="fp16" \
  --gradient_checkpointing \
  --checkpointing_steps=1000 \
  --validation_steps=1000 \
  --report_to="tensorboard" \
  --tracker_project_name="controlnet-colorization" \
  --seed=42

---

License

The model is released under the CreativeML Open RAIL-M license.
You are free to use it for non-commercial and research purposes. Commercial use may require additional permission.

---

Credits

Created by SubMaroon
Trained with compute generously provided by Flanayt Pulsar
Based on the Hugging Face diffusers ControlNet training example
Inspired by lllyasviel's original ControlNet