Manga/Anime Image De-Censoring Pipeline

This workflow is an automated batch image de-censoring pipeline for manga/anime images. It loads censored (mosaic-pixelated) images from a directory, detects mosaic regions using dual detection strategies, generates AI prompts from the image content itself, performs diffusion-based inpainting to remove mosaics, optionally removes backgrounds, and saves the results with structured filenames.

How to Use

1. Setup Input

- Set the Path (PrimitiveString) to your input image folder

- Images are loaded one-at-a-time via Load Image Batch with incremental indexing

2. Select Checkpoint

- Use CheckpointNameSelector to pick your SDXL anime model

- The current default is JANKUTrainedChenkinNoobai_v69.safetensors

3. Ensure LM Studio is Running

- The workflow uses a local LLM nemotron3-nano-4b-uncensored-hauhaucs-aggressive) via LM Studio for prompt generation

- Start LM Studio and load the model before running

4. Run the Workflow

- The pipeline processes automatically: detect mosaic → generate prompt → inpaint → save

5. Output

- Results are saved to a subfolder named after the checkpoint: <save_dir>/<checkpoint_name>/<filename>_<idx>

Workflow Pipeline

Load Image Batch

→ Split Image (RGB + Alpha)

→ Get Mosaic Mask (subgraph with dual detection)

→ Shrink & Blur Mask

→ Is Mosaic Detected?

→ YES: Generate Prompt (WD14 Tagger + LLM)

→ DetailerForEach (1st pass inpainting)

→ ColorMatch (harmonize with original)

→ [Optional: Background Removal (RMBG)]

→ Save Image

Features

Dual Mosaic Detection (Subgraph)

- Primary: ML-based detection using mosaicDetectionAllIn_v40.pt (Ultralytics) + SAM for precise mask refinement (kudos for wildcats Creator Profile | Civitai )

- Fallback: CV-based MosaicDetectionNode with gradient/histogram analysis in HYBRID mode

- Automatically switches between methods based on detection results

Automatic Prompt Generation

- WD14 Tagger wd-v1-4-moat-tagger-v2) tags the input image

- LLM (via LM Studio) cleans tags into diffusion-friendly prompts, removing censorship references

- No manual prompting required — the workflow prompts itself from image content

Smart Inpainting

- First-pass inpainting with DetailerForEach: 1024px guide size, 12 steps, CFG 7, euler_ancestral scheduler, denoise 0.6

- Mask shrink (16px) + gaussian blur (7px) for tight inpaint regions

- INPAINT_ColorMatch ensures seamless blending with the original image

Negative Prompt Anti-Censorship

- Explicitly targets censorship artifacts: censored, mosaic censor, blur, censorship...

- Steers the model away from generating new censorship

NSFW Detailing

- Second-pass refinement using separate detectors for genitals

- 15 steps, CFG 5, denoise 0.4

- Enable via Fast Groups Bypasser if needed

Optional Background Removal

- RMBG-2.0 model removes backgrounds with alpha transparency output

- Controlled via is_alpha_bg variable

Batch Processing

- Processes entire directories sequentially

- Structured output naming with checkpoint and timestamp

Group Bypassing

- Fast Groups Bypasser (rgthree) widget to quickly enable/disable workflow sections

Required Models

| Model | Type | Purpose |

|-------|------|----------|

| JANKUTrainedChenkinNoobai_v69.safetensors | Checkpoint | SDXL anime inpainting |

| sam_vit_b_01ec64.pth | SAM | Precise mask refinement |

| mosaic detection(All-in-one) - v4.0 | Other Detection | Civitai | Ultralytics | ML mosaic detection |

| Anime NSFW Detection / ADetailer All-in-One - v5.0-variant1 | Other Detection | Civitai| Ultralytics | NSFW segmentation (group 2) |

| wd-v1-4-moat-tagger-v2 | Tagger | Anime image tagging |

| HauhauCS/Nemotron3-Nano-4B-Uncensored-HauhauCS-Aggressive · Hugging Face (LM Studio) | LLM | Prompt cleaning |

| 1038lab/ComfyUI-RMBG: A ComfyUI custom node designed for advanced image background removal and object, face, clothes, and fashion segmentation, utilizing multiple models including RMBG-2.0, INSPYRENET, BEN, BEN2, BiRefNet, SDMatte, SAM, SAM2, SAM3 and GroundingDINO. | RMBG | Background removal |

## Groups

1. SDXL — Checkpoint loading & global model broadcasting

2. Inputs — Image batch loading, path/filename setup, seed

3. 1. demosaic to NSFW — Mosaic detection, mask generation, first-pass inpainting

4. 2. NSFW detailing — Second-pass body-part refinement (bypassed by default)

5. Generate prompt from image — WD14 tagging + LLM prompt cleaning

6. Final save — Optional background removal & image saving