Dual Checkpoint Image Generation with Face Detailer & Upscaling

ComfyUI workflow featuring dual checkpoint architecture, multi-LoRA management, and progressive enhancement pipeline

This workflow uses a three-stage processing approach: base generation, face enhancement, and neural upscaling.

🚀 Quick Installation Guide

Required Custom Node Packs

1. ComfyUI-Impact-Pack

   • Provides: FaceDetailer, UltralyticsDetectorProvider

2. ComfyUI-LoraManager

   • Provides: Lora Loader, Debug Metadata, TriggerWord Toggle

3. rgthree-comfy

   • Provides: Fast Groups Muter, Power Prompt - Simple

4. ComfyUI-Custom-Scripts

   • Provides: ShowText|pysssss, CheckpointLoader|pysssss

5. ComfyUI-KJNodes

   • Provides: JoinStringMulti, ImageResizeKJv2

6. ComfyUI-Studio-Nodes (Optional)

   • Provides: AspectRatioImageSize

Required Models

Detection Model:

• bbox/face_yolov8m.pt - Face detection model

• Auto-downloads to: ComfyUI/models/ultralytics/bbox/

Upscale Model:

• 4x-AnimeSharp.pth or 4x-UltraSharp.pth

• Download from: OpenModelDB or Upscale Wiki

• Place in: ComfyUI/models/upscale_models/

Checkpoint Models: (User provided)

• Base checkpoint(s)

• Refinement checkpoint(s)

LoRA Models: (User provided)

• LoRAs as needed for your generation

Installation Methods

Option 1: ComfyUI Manager (Recommended)

1. Install ComfyUI-Manager

2. Load the workflow in ComfyUI

3. Use "Install Missing Custom Nodes" button

4. Restart ComfyUI

Option 2: Manual Installation

cd ComfyUI/custom_nodes

git clone https://github.com/ltdrdata/ComfyUI-Impact-Pack.git
git clone https://github.com/Suzie1/ComfyUI-LoraManager.git
git clone https://github.com/rgthree/rgthree-comfy.git
git clone https://github.com/pythongosssss/ComfyUI-Custom-Scripts.git
git clone https://github.com/kijai/ComfyUI-KJNodes.git
git clone https://github.com/comfyuistudio/ComfyUI-Studio-nodes.git

cd ComfyUI-Impact-Pack
python install.py

Additional Python Dependencies

Some nodes may require additional Python packages:

• Impact Pack: ultralytics, segment-anything, mmdet

• KJNodes: May need numba for some operations

Notes:

• Core ComfyUI nodes are included with base ComfyUI

• Some nodes install their own dependencies on first run

• The workflow will show red/missing nodes if dependencies are missing

🏗️ Workflow Architecture

Three-Stage Processing Pipeline

Stage 1: Base Generation

• Initial image generation

• Dual checkpoint support

• Multi-LoRA management

• Prompt processing and conditioning

Stage 2: Face Enhancement

• Face detection using YOLOv8

• Targeted inpainting and refinement

• Uses secondary checkpoint

• Adaptive denoising

Stage 3: Neural Upscaling

• AI model-based upscaling

• Tile-based processing

• Edge preservation

• Multiple save points with metadata

📦 Main Node Types Used

Model Loading

CheckpointLoader|pysssss

• Loads checkpoint models

• Outputs: MODEL, CLIP, VAE

• Enhanced checkpoint loader with metadata features

VAELoader

• Loads VAE models separately

• Allows VAE selection independent of checkpoint

Prompt Processing

Power Prompt - Simple (rgthree)

• Prompt input and processing

• Supports prompt weighting syntax

• Outputs: CONDITIONING and TEXT

CLIPTextEncode

• Converts text prompts to CLIP embeddings

• Separate nodes for positive and negative prompts

JoinStringMulti

• Combines multiple text strings

• Used for merging trigger words with prompts

ShowText|pysssss

• Displays text output

• Useful for debugging prompts

LoRA Management

Lora Loader (LoraManager)

• Manages multiple LoRA models

• Individual strength controls for each LoRA

• Separate model and CLIP strength settings

• Automatically extracts trigger words

• Toggle system to enable/disable LoRAs

TriggerWord Toggle (LoraManager)

• Manages trigger words from active LoRAs

• Filters based on enabled LoRAs

• Group mode for batch management

Dimension Management

AspectRatioImageSize

• Calculates dimensions for generation

• Preset aspect ratios available

• Ensures VAE-compatible dimensions (divisible by 8)

EmptyLatentImage

• Creates initial latent tensor

• Supports batch generation

Sampling

KSampler

• Core generation node

• Configurable samplers (DPM++, Euler, DDIM, etc.)

• Configurable schedulers (Karras, exponential, simple, etc.)

• Adjustable steps, CFG scale, and denoise strength

Face Detection and Enhancement

UltralyticsDetectorProvider

• Provides YOLOv8 face detection model

• Generates bounding boxes for detected faces

FaceDetailer

• Enhances detected face regions

• Performs targeted inpainting

• Uses separate model for face processing

• Configurable denoise, crop factor, feathering

• Supports SAM model integration

• Processes faces at higher resolution

Image Processing

VAEEncode

• Converts pixel images to latent space

VAEDecode

• Converts latent tensors to pixel images

ImageResizeKJv2

• Resizes images with multiple interpolation methods

• Maintains aspect ratios

• Divisibility enforcement for model compatibility

UpscaleModelLoader

• Loads neural upscaling models (ESRGAN, etc.)

ImageUpscaleWithModel

• Applies neural upscaling to images

• Tile-based processing for large images

Utilities

LazySwitchKJ

• Routes connections based on boolean switch

• Used for conditional workflow paths

WildcardPromptFromString

• Processes wildcard syntax in prompts

Debug Metadata (LoraManager)

• Tracks generation parameters

• Outputs metadata for documentation

SaveImageWithMetaData

• Saves images with embedded metadata

• Configurable file naming and organization

• Multiple instances for different pipeline stages

🔧 Workflow Structure

The workflow uses:

• 2 Checkpoint Loaders - Dual checkpoint architecture

• 2 VAE Loaders - Separate VAE selection

• 2 KSamplers - Base generation and refinement

• 1 LoRA Loader - Multi-LoRA management

• 1 FaceDetailer - Face enhancement

• 2 Upscale nodes - Neural upscaling

• 4 Save nodes - Multiple output points

• 4 Debug Metadata nodes - Parameter tracking

Total: 36 nodes in the workflow

⚙️ Key Features

Dual Checkpoint Support

Load different checkpoint models for base generation and face refinement, allowing specialized models for different stages.

Multi-LoRA Management

LoraManager system allows:

• Loading multiple LoRAs simultaneously

• Individual strength control per LoRA

• Toggle activation without reloading

• Automatic trigger word extraction and filtering

Face Enhancement Pipeline

YOLOv8 detection → FaceDetailer inpainting:

• Automatic face detection

• Higher resolution processing for faces

• Separate model for face refinement

• Configurable enhancement strength

Progressive Enhancement

Three-stage approach:

1. Generate base image

2. Enhance detected faces

3. Upscale final result

Metadata Tracking

Debug Metadata nodes throughout pipeline track:

• Generation parameters

• LoRA configurations

• Model settings

• For reproducibility and documentation

📋 Workflow Groups

The workflow organizes nodes into functional groups:

• Prompt Creation

• Model Loaders - Base Image

• Model Loaders - Face Detail

• Base Image generation and save

• Face Detailer settings and save

• Core Image Upscale and save

• Final Upscale and save

• Subdirectory configuration

• Base Resolution settings

💾 Output Management

Multiple save points capture different stages:

• Post-base generation

• Post-face enhancement

• Post-first upscale

• Post-final upscale

Each save node:

• Embeds metadata

• Customizable file naming

• Subdirectory organization

• Configurable output format

🎯 Usage

1. Load checkpoint models

2. Load LoRA models (optional)

3. Configure prompts (positive and negative)

4. Set generation parameters (steps, CFG, sampler, scheduler)

5. Set resolution via AspectRatioImageSize

6. Configure face enhancement settings

7. Select upscale model

8. Queue and generate

The workflow saves outputs at multiple stages, allowing comparison of results throughout the pipeline.

This workflow provides a complete pipeline from initial generation through face enhancement to final upscaling with comprehensive parameter control and metadata tracking.