ImageStylize V2 | Real-to-Anime & Style Transfer Engine (SDXL)

Transform real-world photographic portraits into stunning, highly stylized anime illustrations without losing character identity or composition. ImageStylize V2 is a highly refined, production-ready SDXL image-to-image workflow built to push the boundaries of artistic style transfers while keeping your generations completely under control.

Whether you are converting photos to vibrant cell-shaded illustrations or shifting between highly stylized artistic mediums, this workflow eliminates the usual headache of distorted faces, mutated hands, and broken compositions.

🚀 Core Features

• IPAdapter FaceID Integration: Built-in face identity embedding loops (using FaceID PlusV2) lock down your subject's likeness. It bridges the gap between a 2D style transformation and real-world facial proportions, ensuring the character actually looks like the source image.

• Dual-Layer ControlNet Stack: Combines spatial depth analysis and clean edge boundaries to ensure structural integrity. Your character's exact pose, clothing folds, and limb placement are perfectly preserved.

• Smart Face & Hand Auto-Detailers: Integrated Impact Pack pipelines automatically run secondary, localized inpainting passes on faces and hands. This eliminates blurred expressions and cleans up unstable AI hand geometry on the fly.

• High-Res Polish Stack: Features an optimized upscaling route paired with a micro-dose image noise generator. This acts as a subtle film or cel grain, completely removing ugly digital color banding and preparing your images for crisp uploads.

📦 Required Models & Directory Setup Guide

To run this workflow successfully, download the following models and place them into their respective ComfyUI directories. (Note: Most of these can also be easily searched and downloaded directly via the ComfyUI Manager's "Model Manager" tool).

1. Base Checkpoint & LoRAs

• SDXL Base Checkpoint: Sakura Dreams (or your preferred SDXL Stylized Base)

  • 📂 Directory: ComfyUI/models/checkpoints/

• IP-Adapter FaceID LoRA: ip-adapter-faceid-plusv2_sdxl_lora.safetensors

  • 📂 Directory: ComfyUI/models/loras/

2. ControlNets & Adapters

• ControlNet Depth: controlnet-depth-sdxl-10.safetensors

  • 📂 Directory: ComfyUI/models/controlnet/

• ControlNet Canny/Lineart: controlnet-canny-sdxl-10.safetensors

  • 📂 Directory: ComfyUI/models/controlnet/

• IP-Adapter Model: ip-adapter-faceid-plusv2_sdxl.bin

  • 📂 Directory: ComfyUI/models/ipadapter/

• CLIP Vision Encoder: CLIP-ViT-bigG-14-laion2B-39B-b160k.safetensors

  • 📂 Directory: ComfyUI/models/clip_vision/

3. YOLO Detectors (Impact Pack)

• YOLO Face Detector: face_yolov8m.pt

  • 📂 Directory: ComfyUI/models/ultralytics/bbox/

• YOLO Hand Detector: hand_yolov8s.pt

  • 📂 Directory: ComfyUI/models/ultralytics/bbox/

4. Upscalers

• Pre-Attention Face Isolation Upscaler: 4x-UltraSharp.pth

  • 📂 Directory: ComfyUI/models/upscale_models/

• Final Artwork Upscaler: RealESRGAN_x4plus_anime_6B.pth

  • 📂 Directory: ComfyUI/models/upscale_models/

💡 Crucial Step for IP-Adapter FaceID: FaceID requires the insightface Python library. If you encounter a ValueError or compilation error on boot, run the following command in your root directory terminal to compile it against your environment:

.\python_embeded\python.exe -m pip install insightface --no-binary insightface --no-build-isolation --force-reinstall

🎛️ Recommended Starter Settings (Real ➔ Anime)

To get the absolute most out of this styler right out of the box, load up an SDXL anime/stylized checkpoint and dial in these baselines:

• KSampler Denoise: 0.60 - 0.70 (The sweet spot for giving the model enough power to redraw textures without losing the original photo layout)

• CFG Scale: 3.5 - 5.0 (Keeps colors vibrant without clipping lines or burning the image)

• Sampler/Scheduler: dpmpp_2m / karras

• ControlNet Weights: Depth (0.65 strength, ends at 0.60) | Edge/Lineart (0.35 strength, ends at 0.75)

• Face / Hand Detailer Denoise: Face (0.50) | Hand (0.40)

📦 Required Custom Nodes

Make sure you have updated these suites via the ComfyUI Manager before running:

1. ComfyUI-Easy-Use (For the optimized modular loader loops)

2. ComfyUI-Impact-Pack (For the segmentation, detection, and detailer loops)

3. ComfyUI-ControlNet-Auxiliary-Preprocessors (For clean depth and boundary tracking)

4. ComfyUI_IPAdapter_plus (For facial vector integration)

5. ComfyUI-KJNodes (For final noise/grain polish)