Dual-Checkpoint TIPO-Enhanced SDXL Image Generation — v1.0

Workflow Structure & Dependencies

Required Custom Nodes

Install these custom nodes via ComfyUI Manager :

z-tipo-extension : For TIPO prompt enhancement

comfyui-prompt-control : A1111-style prompt scheduling

ComfyUI-Impact-Pack : FaceDetailer and detection systems

ComfyUI_UltimateSDUpscale : Professional upscaling

ComfyUI_Fill-Nodes : Random number generation

comfyui_controlnet_aux : ControlNet preprocessing

ComfyUI-Easy-Use : Workflow automation helpers

Essential Models

Primary Models :

SDXL-based checkpoint (IllustriousXL recommended)

Secondary SDXL checkpoint for high-res fix

TIPO-500M model for prompt enhancement

Supporting Models :

SAM models for segmentation

YOLO detection models (face, hand, hair)

4x upscaling models (UltraSharp recommended)

ControlNet models (LineArt, Pose)

Artistic Freedom Through Dual Checkpoints

The Approach

By utilizing two different SDXL-based models in sequence, you can:

Initial Generation : Use a specialized checkpoint (e.g., IllustriousXL) for its extensive knowledge of anime artists and character consistency

High-Resolution Refinement : Apply a second checkpoint (e.g., realistic SDXL) to enhance details, lighting, and overall realism

Why This Matters

IllustriousXL brings unparalleled anime artist knowledge and character consistency:

Trained on vast anime datasets with superior character anatomy

Eliminates common hand/foot artifacts present in other models

Extensive pose and composition capabilities

Realistic SDXL Models provide:

Advanced lighting and texture understanding

Photorealistic detail enhancement

Improved background and environmental elements

A lot more Artist knowledge

The Combination results in:

Anime characters with realistic lighting and textures

Consistent character features with enhanced detail quality

Artistic styles impossible with single-model approaches

Technical Implementation

TIPO Integration

TIPO (Text to Image with text Presampling for Prompt Optimization) automatically enhances your prompts:

Input: "1girl, outdoors, sunset" TIPO Output: "1girl, outdoors, sunset, masterpiece, best quality, amazing quality, very aesthetic, ultra-detailed, highly detailed, realistic, beautiful lighting, golden hour, warm colors, detailed background"

Configuration :

Model: KBlueLeaf/TIPO-500M-ft

Operation: short_to_tag_to_long

Temperature: 1.0, Top-p: 0.95

Prompt Control Features

The workflow utilizes advanced prompt control enabling:

A1111-style syntax : (emphasis:1.2) , [negative] , {choices|alternatives}

LoRA scheduling : with dynamic weights

Prompt filtering : Conditional elements based on generation parameters

Regional prompting : Area-specific styling and control

Automation & Randomization

Dynamic Parameter Control :

Aspect Ratios : Randomly selected from portrait, landscape, and square formats

CFG Scale : Range-based randomization (3.0-8.0) for varied artistic interpretation

LoRA Selection : Automated loading from categorized folders with weight randomization

Seed Management : Increment mode for easy iteration and comparison

Professional Enhancement Pipeline

Multi-Stage Detailing

Face Enhancement : Primary face detection and refinement using specialized models

Hand Detailing : Targeted hand improvement with dedicated YOLO models

Hair Refinement : Advanced hair texture and detail enhancement

Final Polish : Comprehensive detail pass with adjustable parameters

Ultimate SD Upscale Integration

Professional Upscaling Features :

Tile-based Processing : Handles large images without memory issues

Seamless Blending : Eliminates tile boundaries through advanced algorithms

Multiple Passes : Iterative refinement for maximum quality

Configurable Denoise : Balance between detail addition and original preservation

Setup Instructions

1. Installation

Install ComfyUI Manager cd ComfyUI/custom_nodes git clone https://github.com/ltdrdata/ComfyUI-Manager.git # Restart ComfyUI and use Manager to install required nodes

2. Model Preparation

Download Required Models :

Place SDXL checkpoints in models/checkpoints/

Download TIPO-500M from HuggingFace (should be done by the TIPO node)

Install detection models via ComfyUI Manager

Configure upscaling models in models/upscale_models/

3. Workflow Loading

Download the provided workflow JSON

Import via ComfyUI interface or drag-and-drop

Install missing nodes when prompted

Configure model paths and preferences

Usage Guidelines:

Basic Operation

Set Primary Checkpoint : Choose your main artistic model (IllustriousXL recommended)

Configure Secondary Checkpoint : Select refinement model for high-fix pass

Input Base Prompt : Simple description that TIPO will enhance

Adjust Parameters : Set quality preferences and generation count

Queue Generation : Let automation handle the rest

Advanced Configuration

For Maximum Artistic Control :

Modify LoRA categories and weights

Adjust detailing passes and strengths

Configure ControlNet inputs for pose/composition control

Fine-tune upscaling parameters for output quality (make sure you can divide the resolution with 64)

Best Practices

Start Simple : Begin with basic settings before adding complexity

Test Incrementally : Enable features one at a time to isolate issues

Monitor Resources : Watch GPU memory usage during long generations

Save Configurations : Use ComfyUI's workflow saving for reproducible results

Object Detector · Hugging Face