<h1 id="mangaanime-image-de-censoring-pipeline">Manga/Anime Image De-Censoring Pipeline</h1><p>This workflow is an <strong>automated batch image de-censoring pipeline</strong> 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.</p><h2 id="how-to-use"> How to Use</h2><h3 id="1.-setup-input"> 1. Setup Input</h3><p>- Set the <strong>Path</strong> (PrimitiveString) to your input image folder</p><p>- Images are loaded one-at-a-time via <code>Load Image Batch</code> with incremental indexing</p><h3 id="2.-select-checkpoint">2. Select Checkpoint</h3><p>- Use <code>CheckpointNameSelector</code> to pick your SDXL anime model</p><p>- The current default is <code>JANKUTrainedChenkinNoobai_v69.safetensors</code></p><h3 id="3.-ensure-lm-studio-is-running">3. Ensure LM Studio is Running</h3><p>- The workflow uses a local LLM <code>nemotron3-nano-4b-uncensored-hauhaucs-aggressive</code>) via LM Studio for prompt generation</p><p>- Start LM Studio and load the model before running</p><h3 id="4.-run-the-workflow">4. Run the Workflow</h3><p>- The pipeline processes automatically: detect mosaic → generate prompt → inpaint → save</p><h3 id="5.-output">5. Output</h3><p>- Results are saved to a subfolder named after the checkpoint: <code>&lt;save_dir&gt;/&lt;checkpoint_name&gt;/&lt;filename&gt;_&lt;idx&gt;</code></p><h2 id="workflow-pipeline">Workflow Pipeline</h2><p>Load Image Batch</p><p>  → Split Image (RGB + Alpha)</p><p>    → Get Mosaic Mask (subgraph with dual detection)</p><p>      → Shrink &amp; Blur Mask</p><p>        → Is Mosaic Detected?</p><p>          → YES: Generate Prompt (WD14 Tagger + LLM)</p><p>          → DetailerForEach (1st pass inpainting)</p><p>            → ColorMatch (harmonize with original)</p><p>              → [Optional: Background Removal (RMBG)]</p><p>                → Save Image</p><h2 id="features">Features</h2><h3 id="dual-mosaic-detection-(subgraph)">Dual Mosaic Detection (Subgraph)</h3><p>- <strong>Primary</strong>: ML-based detection using <code>mosaicDetectionAllIn_v40.pt</code> (Ultralytics) + SAM for precise mask refinement (kudos for <a target="_blank" rel="ugc" href="https://civarchive.com/user/wildcats">wildcats Creator Profile | Civitai</a> )</p><p>- <strong>Fallback</strong>: CV-based <code>MosaicDetectionNode</code> with gradient/histogram analysis in HYBRID mode</p><p>- Automatically switches between methods based on detection results</p><h3 id="automatic-prompt-generation">Automatic Prompt Generation</h3><p>- <strong>WD14 Tagger</strong> <code>wd-v1-4-moat-tagger-v2</code>) tags the input image</p><p>- <strong>LLM</strong> (via LM Studio) cleans tags into diffusion-friendly prompts, removing censorship references</p><p>- No manual prompting required — the workflow prompts itself from image content</p><h3 id="smart-inpainting">Smart Inpainting</h3><p>- First-pass inpainting with <code>DetailerForEach</code>: 1024px guide size, 12 steps, CFG 7, euler_ancestral scheduler, denoise 0.6</p><p>- Mask shrink (16px) + gaussian blur (7px) for tight inpaint regions</p><p>- <code>INPAINT_ColorMatch</code> ensures seamless blending with the original image</p><h3 id="negative-prompt-anti-censorship">Negative Prompt Anti-Censorship</h3><p>- Explicitly targets censorship artifacts: <code>censored, mosaic censor, blur, censorship...</code></p><p>- Steers the model away from generating new censorship</p><h3 id="nsfw-detailing">NSFW Detailing</h3><p>- Second-pass refinement using separate detectors for genitals</p><p>- 15 steps, CFG 5, denoise 0.4</p><p>- Enable via <code>Fast Groups Bypasser</code> if needed</p><h3 id="optional-background-removal">Optional Background Removal</h3><p>- <code>RMBG-2.0</code> model removes backgrounds with alpha transparency output</p><p>- Controlled via <code>is_alpha_bg</code> variable</p><h3 id="batch-processing">Batch Processing</h3><p>- Processes entire directories sequentially</p><p>- Structured output naming with checkpoint and timestamp</p><h3 id="group-bypassing">Group Bypassing</h3><p>- <code>Fast Groups Bypasser (rgthree)</code> widget to quickly enable/disable workflow sections</p><h3 id="required-models">Required Models</h3><p>| Model | Type | Purpose |</p><p>|-------|------|----------|</p><p>| <code>JANKUTrainedChenkinNoobai_v69.safetensors</code> | Checkpoint | SDXL anime inpainting |</p><p>| <code>sam_vit_b_01ec64.pth</code> | SAM | Precise mask refinement |</p><p>| <a target="_blank" rel="ugc" href="https://civarchive.com/models/1308310?modelVersionId=2480170">mosaic detection(All-in-one) - v4.0 | Other Detection | Civitai</a> | Ultralytics | ML mosaic detection |</p><p>| <a target="_blank" rel="ugc" href="https://civarchive.com/models/1313556/anime-nsfw-detection-adetailer-all-in-one">Anime NSFW Detection / ADetailer All-in-One - v5.0-variant1 | Other Detection | Civitai</a>| Ultralytics | NSFW segmentation (group 2) |</p><p>| <code>wd-v1-4-moat-tagger-v2</code> | Tagger | Anime image tagging |</p><p>| <a target="_blank" rel="ugc" href="https://huggingface.co/HauhauCS/Nemotron3-Nano-4B-Uncensored-HauhauCS-Aggressive">HauhauCS/Nemotron3-Nano-4B-Uncensored-HauhauCS-Aggressive · Hugging Face </a>(LM Studio) | LLM | Prompt cleaning |</p><p>| <a target="_blank" rel="ugc" href="https://github.com/1038lab/ComfyUI-RMBG">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. </a>| RMBG | Background removal |</p><p>## Groups</p><p>1. <strong>SDXL</strong> — Checkpoint loading &amp; global model broadcasting</p><p>2. <strong>Inputs</strong> — Image batch loading, path/filename setup, seed</p><p>3. <strong>1. demosaic to NSFW</strong> — Mosaic detection, mask generation, first-pass inpainting</p><p>4. <strong>2. NSFW detailing</strong> — Second-pass body-part refinement (bypassed by default)</p><p>5. <strong>Generate prompt from image</strong> — WD14 tagging + LLM prompt cleaning</p><p>6. <strong>Final save</strong> — Optional background removal &amp; image saving</p>

Manga/Anime Image De-Censoring Pipeline

<p>Initial version. May have issues and mistakes</p>

Files with SHA256: c3ae9c9ef4d0c7f196fdeef4493a042bb1ec33b1fe1e28d6a52b42234c17a2e2