<h1 id="dual-checkpoint-image-generation-with-face-detailer-and-upscaling">Dual Checkpoint Image Generation with Face Detailer &amp; Upscaling</h1>ComfyUI workflow featuring dual checkpoint architecture, multi-LoRA management, and progressive enhancement pipelineThis workflow uses a three-stage processing approach: base generation, face enhancement, and neural upscaling.<hr /><h2 id="quick-installation-guide">🚀 Quick Installation Guide</h2><h3 id="required-custom-node-packs">Required Custom Node Packs</h3><ol><li><a target="_blank" rel="ugc" href="https://github.com/ltdrdata/ComfyUI-Impact-Pack">ComfyUI-Impact-Pack</a><ul><li>Provides: FaceDetailer, UltralyticsDetectorProvider</li></ul></li><li><a target="_blank" rel="ugc" href="https://github.com/Suzie1/ComfyUI-LoraManager">ComfyUI-LoraManager</a><ul><li>Provides: Lora Loader, Debug Metadata, TriggerWord Toggle</li></ul></li><li><a target="_blank" rel="ugc" href="https://github.com/rgthree/rgthree-comfy">rgthree-comfy</a><ul><li>Provides: Fast Groups Muter, Power Prompt - Simple</li></ul></li><li><a target="_blank" rel="ugc" href="https://github.com/pythongosssss/ComfyUI-Custom-Scripts">ComfyUI-Custom-Scripts</a><ul><li>Provides: ShowText|pysssss, CheckpointLoader|pysssss</li></ul></li><li><a target="_blank" rel="ugc" href="https://github.com/kijai/ComfyUI-KJNodes">ComfyUI-KJNodes</a><ul><li>Provides: JoinStringMulti, ImageResizeKJv2</li></ul></li><li><a target="_blank" rel="ugc" href="https://github.com/comfyuistudio/ComfyUI-Studio-nodes">ComfyUI-Studio-Nodes</a> (Optional)<ul><li>Provides: AspectRatioImageSize</li></ul></li></ol><h3 id="required-models">Required Models</h3>Detection Model:<ul><li><code>bbox/face_yolov8m.pt</code> - Face detection model</li><li>Auto-downloads to: <code>ComfyUI/models/ultralytics/bbox/</code></li></ul>Upscale Model:<ul><li><code>4x-AnimeSharp.pth</code> or <code>4x-UltraSharp.pth</code></li><li>Download from: <a target="_blank" rel="ugc" href="https://openmodeldb.info/">OpenModelDB</a> or <a target="_blank" rel="ugc" href="https://upscale.wiki/">Upscale Wiki</a></li><li>Place in: <code>ComfyUI/models/upscale_models/</code></li></ul>Checkpoint Models: (User provided)<ul><li>Base checkpoint(s)</li><li>Refinement checkpoint(s)</li></ul>LoRA Models: (User provided)<ul><li>LoRAs as needed for your generation</li></ul><h3 id="installation-methods">Installation Methods</h3>Option 1: ComfyUI Manager (Recommended)<ol><li>Install <a target="_blank" rel="ugc" href="https://github.com/ltdrdata/ComfyUI-Manager">ComfyUI-Manager</a></li><li>Load the workflow in ComfyUI</li><li>Use "Install Missing Custom Nodes" button</li><li>Restart ComfyUI</li></ol>Option 2: Manual Installation<pre><code>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
</code></pre><h3 id="additional-python-dependencies">Additional Python Dependencies</h3>Some nodes may require additional Python packages:<ul><li>Impact Pack: ultralytics, segment-anything, mmdet</li><li>KJNodes: May need numba for some operations</li></ul>Notes:<ul><li>Core ComfyUI nodes are included with base ComfyUI</li><li>Some nodes install their own dependencies on first run</li><li>The workflow will show red/missing nodes if dependencies are missing</li></ul><hr /><h2 id="workflow-architecture">🏗️ Workflow Architecture</h2><h3 id="three-stage-processing-pipeline">Three-Stage Processing Pipeline</h3>Stage 1: Base Generation<ul><li>Initial image generation</li><li>Dual checkpoint support</li><li>Multi-LoRA management</li><li>Prompt processing and conditioning</li></ul>Stage 2: Face Enhancement<ul><li>Face detection using YOLOv8</li><li>Targeted inpainting and refinement</li><li>Uses secondary checkpoint</li><li>Adaptive denoising</li></ul>Stage 3: Neural Upscaling<ul><li>AI model-based upscaling</li><li>Tile-based processing</li><li>Edge preservation</li><li>Multiple save points with metadata</li></ul><hr /><h2 id="main-node-types-used">📦 Main Node Types Used</h2><h3 id="model-loading">Model Loading</h3>CheckpointLoader|pysssss<ul><li>Loads checkpoint models</li><li>Outputs: MODEL, CLIP, VAE</li><li>Enhanced checkpoint loader with metadata features</li></ul>VAELoader<ul><li>Loads VAE models separately</li><li>Allows VAE selection independent of checkpoint</li></ul><h3 id="prompt-processing">Prompt Processing</h3>Power Prompt - Simple (rgthree)<ul><li>Prompt input and processing</li><li>Supports prompt weighting syntax</li><li>Outputs: CONDITIONING and TEXT</li></ul>CLIPTextEncode<ul><li>Converts text prompts to CLIP embeddings</li><li>Separate nodes for positive and negative prompts</li></ul>JoinStringMulti<ul><li>Combines multiple text strings</li><li>Used for merging trigger words with prompts</li></ul>ShowText|pysssss<ul><li>Displays text output</li><li>Useful for debugging prompts</li></ul><h3 id="lora-management">LoRA Management</h3>Lora Loader (LoraManager)<ul><li>Manages multiple LoRA models</li><li>Individual strength controls for each LoRA</li><li>Separate model and CLIP strength settings</li><li>Automatically extracts trigger words</li><li>Toggle system to enable/disable LoRAs</li></ul>TriggerWord Toggle (LoraManager)<ul><li>Manages trigger words from active LoRAs</li><li>Filters based on enabled LoRAs</li><li>Group mode for batch management</li></ul><h3 id="dimension-management">Dimension Management</h3>AspectRatioImageSize<ul><li>Calculates dimensions for generation</li><li>Preset aspect ratios available</li><li>Ensures VAE-compatible dimensions (divisible by 8)</li></ul>EmptyLatentImage<ul><li>Creates initial latent tensor</li><li>Supports batch generation</li></ul><h3 id="sampling">Sampling</h3>KSampler<ul><li>Core generation node</li><li>Configurable samplers (DPM++, Euler, DDIM, etc.)</li><li>Configurable schedulers (Karras, exponential, simple, etc.)</li><li>Adjustable steps, CFG scale, and denoise strength</li></ul><h3 id="face-detection-and-enhancement">Face Detection and Enhancement</h3>UltralyticsDetectorProvider<ul><li>Provides YOLOv8 face detection model</li><li>Generates bounding boxes for detected faces</li></ul>FaceDetailer<ul><li>Enhances detected face regions</li><li>Performs targeted inpainting</li><li>Uses separate model for face processing</li><li>Configurable denoise, crop factor, feathering</li><li>Supports SAM model integration</li><li>Processes faces at higher resolution</li></ul><h3 id="image-processing">Image Processing</h3>VAEEncode<ul><li>Converts pixel images to latent space</li></ul>VAEDecode<ul><li>Converts latent tensors to pixel images</li></ul>ImageResizeKJv2<ul><li>Resizes images with multiple interpolation methods</li><li>Maintains aspect ratios</li><li>Divisibility enforcement for model compatibility</li></ul>UpscaleModelLoader<ul><li>Loads neural upscaling models (ESRGAN, etc.)</li></ul>ImageUpscaleWithModel<ul><li>Applies neural upscaling to images</li><li>Tile-based processing for large images</li></ul><h3 id="utilities">Utilities</h3>LazySwitchKJ<ul><li>Routes connections based on boolean switch</li><li>Used for conditional workflow paths</li></ul>WildcardPromptFromString<ul><li>Processes wildcard syntax in prompts</li></ul>Debug Metadata (LoraManager)<ul><li>Tracks generation parameters</li><li>Outputs metadata for documentation</li></ul>SaveImageWithMetaData<ul><li>Saves images with embedded metadata</li><li>Configurable file naming and organization</li><li>Multiple instances for different pipeline stages</li></ul><hr /><h2 id="workflow-structure">🔧 Workflow Structure</h2>The workflow uses:<ul><li>2 Checkpoint Loaders - Dual checkpoint architecture</li><li>2 VAE Loaders - Separate VAE selection</li><li>2 KSamplers - Base generation and refinement</li><li>1 LoRA Loader - Multi-LoRA management</li><li>1 FaceDetailer - Face enhancement</li><li>2 Upscale nodes - Neural upscaling</li><li>4 Save nodes - Multiple output points</li><li>4 Debug Metadata nodes - Parameter tracking</li></ul>Total: 36 nodes in the workflow<hr /><h2 id="key-features">⚙️ Key Features</h2><h3 id="dual-checkpoint-support">Dual Checkpoint Support</h3>Load different checkpoint models for base generation and face refinement, allowing specialized models for different stages.<h3 id="multi-lora-management">Multi-LoRA Management</h3>LoraManager system allows:<ul><li>Loading multiple LoRAs simultaneously</li><li>Individual strength control per LoRA</li><li>Toggle activation without reloading</li><li>Automatic trigger word extraction and filtering</li></ul><h3 id="face-enhancement-pipeline">Face Enhancement Pipeline</h3>YOLOv8 detection → FaceDetailer inpainting:<ul><li>Automatic face detection</li><li>Higher resolution processing for faces</li><li>Separate model for face refinement</li><li>Configurable enhancement strength</li></ul><h3 id="progressive-enhancement">Progressive Enhancement</h3>Three-stage approach:<ol><li>Generate base image</li><li>Enhance detected faces</li><li>Upscale final result</li></ol><h3 id="metadata-tracking">Metadata Tracking</h3>Debug Metadata nodes throughout pipeline track:<ul><li>Generation parameters</li><li>LoRA configurations</li><li>Model settings</li><li>For reproducibility and documentation</li></ul><hr /><h2 id="workflow-groups">📋 Workflow Groups</h2>The workflow organizes nodes into functional groups:<ul><li>Prompt Creation</li><li>Model Loaders - Base Image</li><li>Model Loaders - Face Detail</li><li>Base Image generation and save</li><li>Face Detailer settings and save</li><li>Core Image Upscale and save</li><li>Final Upscale and save</li><li>Subdirectory configuration</li><li>Base Resolution settings</li></ul><hr /><h2 id="output-management">💾 Output Management</h2>Multiple save points capture different stages:<ul><li>Post-base generation</li><li>Post-face enhancement</li><li>Post-first upscale</li><li>Post-final upscale</li></ul>Each save node:<ul><li>Embeds metadata</li><li>Customizable file naming</li><li>Subdirectory organization</li><li>Configurable output format</li></ul><hr /><h2 id="usage">🎯 Usage</h2><ol><li>Load checkpoint models</li><li>Load LoRA models (optional)</li><li>Configure prompts (positive and negative)</li><li>Set generation parameters (steps, CFG, sampler, scheduler)</li><li>Set resolution via AspectRatioImageSize</li><li>Configure face enhancement settings</li><li>Select upscale model</li><li>Queue and generate</li></ol>The workflow saves outputs at multiple stages, allowing comparison of results throughout the pipeline.<hr />This workflow provides a complete pipeline from initial generation through face enhancement to final upscaling with comprehensive parameter control and metadata tracking.

Dual Checkpoint LoraManager Studio Workflow with FaceDetailer + Upscale

This workflow works well with SDXL, Pony, Illustrious and many other base models.

Files with SHA256: 3ab2a3cd3abb72b831ec9f236b1a0abcb9c1591d312fa04dccaec0cc4e8bb487