SDXL 4 Phase Super Refiner V1

DarthRidonkulousDarthRidonkulous
2
8
工具
4-Phase Super Refiner V1

详情

下载文件

模型描述

ComfyUI SDXL 4-Phase Super Refiner V1

Ultimate Multi-Phase SDXL Image Generation Workflow

Author: DarthRidonkulous
Version: 1.0
Release Date: November 26, 2025
Status: Ready to Use

🌟 Overview

The SDXL 4-Phase Super Refiner V1 is a creative powerhouse that lets you build images through four distinct transformation phases, each one giving you complete artistic control to shape, refine, and evolve your vision. Think of it as painting in layers—start with your concept, then progressively add detail, change direction, or completely transform the style as you go.

What Makes This Workflow Special

  • Four Creative Transformation Phases - Each phase can take your image in a completely different direction

  • Mix & Match Models - Use different checkpoints per phase for wild hybrid styles

  • Independent LoRA Control - Load up 12+ LoRAs and cherry-pick which effects fire at each phase

  • Style Evolution - Start photorealistic, end up anime. Or vice versa. Or something in between.

  • Experiment Freely - 12+ save points let you see every phase of transformation

  • Triple FaceDetailer - Perfect faces at multiple phases, or let them evolve naturally

  • Creative Chaos Mode - Randomize everything and generate hundreds of variations

📦 Installation Guide

System Requirements

Minimum Requirements:

  • ComfyUI (latest version recommended)

  • NVIDIA GPU with 12GB+ VRAM (16GB+ recommended for full features)

  • 32GB+ system RAM (64GB recommended for large batches)

  • Python 3.10+ (usually included with ComfyUI)

  • SSD storage for faster workflow loading

Recommended Requirements:

  • NVIDIA RTX 4090 or equivalent

  • 64GB+ system RAM

  • NVMe SSD storage

  • ComfyUI Manager installed

Required Custom Node Packs

This workflow requires the following custom node packs (install via ComfyUI Manager):

  • ComfyUI-Impact-Pack - FaceDetailer and detection nodes

  • ComfyUI-LoRA-Manager - Visual LoRA selection interface

  • ComfyUI-Easy-Use - Memory management utilities

  • rgthree's ComfyUI Nodes - Power Prompt functionality

  • pysssss Custom Scripts - ShowText and ReroutePrimitive nodes

  • WAS Node Suite - Various utility nodes

  • kjNodes - Image processing utilities

Custom Nodes Installation

This workflow includes custom controller nodes that must be installed:

Step 1: Locate Custom Nodes Folder

  • Windows: ComfyUI_windows_portable\ComfyUI\custom_nodes\

  • Linux/Mac: ComfyUI/custom_nodes/

Step 2: Install Required Custom Nodes

Copy these files to your custom_nodes folder (available in the project files):

  • sdxl_image_settings_controller_4phase.py - Centralized 4-phase settings management

  • multi_phase_prompt_manager_4phase_v2.py - Advanced 4-phase prompt control with positive/negative prompts

  • vae_selector.py - VAE selection and management

Step 3: Restart ComfyUI

  • Close ComfyUI completely

  • Restart the application

  • Wait for full startup

Step 4: Verify Installation

  • Right-click on canvas to open node menu

  • Search for "SDXL Image Settings Controller (4 Phase)"

  • Search for "Multi-Phase Prompt Manager (4 Phase) V2"

  • Search for "VAE Selector"

  • If all three appear, installation successful!

Step 5: Load the Workflow

  1. Download SDXL_4_Phase_Super_Refiner_V1.json

  2. Open ComfyUI interface

  3. Click "Load" or drag JSON file onto canvas

  4. Workflow should load with all nodes visible

  5. Check for any missing node warnings

What These Custom Nodes Do

SDXL Image Settings Controller (4 Phase)

This is your central control hub for the entire 4-phase workflow. Instead of having settings scattered across multiple nodes, everything is in one place:

  • Resolution Control: 60+ resolution presets including:

    • Square [1:1]: 512² to 1792²

    • Portrait [2:3, 3:4, 9:16]: Multiple aspect ratios

    • Landscape [3:2, 4:3, 16:9]: Multiple aspect ratios

    • Ultra-Wide [21:9, 32:9]: For cinematic content

    • Cinematic [2.39:1]: Anamorphic widescreen

  • Seed Management: Four modes (fixed, randomize, increment, decrement)

  • Batch Size: Generate 1-4096 images at once

  • Subdirectory Organization: Keep your outputs organized by project

  • Per-Phase KSampler Settings: Configure all 5 settings (steps, CFG, sampler, scheduler, denoise) for ALL FOUR phases

  • Automatic Latent Generation: Creates the empty latent at your chosen resolution with correct batch size

  • Metadata Output: Complete summary of all settings for debugging

Outputs (27 total):

  • LATENT (empty latent image)

  • seed, subdirectory, width, height, batch_size

  • All Phase 1 KSampler settings (5 outputs)

  • All Phase 2 KSampler settings (5 outputs)

  • All Phase 3 KSampler settings (5 outputs)

  • All Phase 4 KSampler settings (5 outputs)

  • metadata (text summary)

This node replaces what would normally be 5+ separate nodes and makes it easy to adjust everything from a single location.

Multi-Phase Prompt Manager (4 Phase) V2

This handles BOTH positive and negative prompts across all four phases with independent control:

  • Phase 1: Base positive and negative prompts that other phases can reference

  • Per-Phase Trigger Selection: Each phase (1-4) has its own "selected_triggers" input for LoRA trigger words

  • Five Modes Per Phase (both positive and negative):

    • "Use Base (Phase 1)" - Use Phase 1 prompt

    • "Use Previous Phase" - Use the previous phase's prompt

    • "Unique Prompt" - Completely different prompt

    • "Append to Base" - Add to Phase 1 prompt (default)

    • "Append to Previous Phase" - Add to previous phase

  • Independent Positive AND Negative: Each phase controls both positive and negative prompts separately

  • Segs_Labels Support: Optional delimiter for each phase (wildcard integration)

  • Metadata Output: Debug summary showing which modes each phase is using

The key feature: Each phase gets its OWN trigger words from LoRA Manager. Phase 1 triggers don't affect Phase 2, Phase 2 doesn't affect Phase 3, etc. This is true per-phase LoRA control with no cross-contamination.

VAE Selector

This node lets you choose between using the VAE baked into your checkpoint or loading an external VAE file:

  • Toggle Switch: Simple on/off to choose between baked or external VAE

  • Baked VAE Mode (default): Uses the VAE that came with your checkpoint (passthrough)

  • External VAE Mode: Loads a separate VAE file from your VAE folder

  • Quick Testing: Easily compare baked vs external VAE quality without rewiring

This is useful when you want to test if an external VAE (like sdxl_vae.safetensors) produces better results than the VAE baked into your checkpoint, or when certain checkpoints need specific VAEs to work properly.

⚡ Centralized Control and Review Location

The workflow is organized with centralized frames for easy access and monitoring:

"Step 1- Phases, Models and Settings" Frame (Control Center):

  • SDXL Image Settings Controller (4 Phase) - All resolution, seed, batch, and KSampler settings for all 4 phases

  • Phase 1/2/3/4 Model Loaders - Checkpoint selection for each phase

  • Multi-Phase Prompt Manager (4 Phase) V2 - All prompts and trigger words for all 4 phases

  • VAE Selector - VAE selection

"Results and Settings" Frames (Review Areas):

There are two "Results and Settings" frames in the workflow - one for Phases 1-2 and one for Phases 3-4:

  • Preview Images - Large image preview nodes showing final image output from each phase

    • Phase 1 FaceDetailer preview

    • Phase 2 FaceDetailer preview

    • Phase 3 FaceDetailer preview

    • Phase 4 FaceDetailer preview

  • ShowText Nodes - Display complete settings for verification:

    • SDXL Image Settings Controller metadata (resolution, seed, all KSampler settings)

    • Multi-Phase Prompt Manager configuration (shows which modes each phase is using)

    • Positive and negative prompts for each phase

    • LoRA trigger words being applied

  • Quick Visual Comparison - See all phase outputs side-by-side without scrolling

What This Means:

You don't need to scroll through the massive workflow to check settings or see results. Everything is organized:

  1. Make changes in the "Step 1" control frame

  2. Verify settings in the ShowText nodes in "Results and Settings"

  3. Review outputs in the large preview images

  4. Compare phases side-by-side to see the evolution

This centralized approach means you can adjust settings, verify configuration, and review results all from one area without hunting through 700+ settings.

🎛️ Bypass Control - Use Only What You Need

You don't have to run all 4 phases every time. The workflow is designed so you can bypass entire phases and optional features:

Bypass Phases:

  • Use the Fast Bypass Node to bypass each phase and the model loader for that phase. You can also bypass save nodes for images that you don’t want to save or see.

  • Run Phase 1 only for quick prompt testing

  • Run Phases 1-2 only for faster iterations

  • Enable all 4 phases for final renders

Bypass FaceDetailers and Upscalers:

  • Phase 1 through 4 Face Detailers can be bypassed. Enable only the ones you need, or none at all

Bypass Save Nodes:

  • 12+ save points throughout the workflow

  • Disable intermediate saves you don't need

  • Keep only Phase 4 final output if you want

  • Or keep all saves to see complete evolution

Bypass Upscaling:

  • Upscaling is optional

  • Disable if you don't need 4x resolution in any Phase

  • Saves significant processing time

Why This Matters:

  • Speed: Bypass phases for faster testing

  • VRAM: Disable features if you're running low on memory

  • Disk Space: Turn off save nodes you don't need

  • Workflow Flexibility: Use as much or as little as you want

The workflow runs whatever you leave enabled. Everything else is skipped.

🚀 Quick Start Guide

Basic Workflow Setup

Step 1: Choose Which Phases to Run

Locate the Fast Bypass node (rgthree Fast Groups Bypasser):

  • Controls: Phase 2, Phase 3, Phase 4, FaceDetailers, and Save nodes

  • Enable phases you want to use (checkboxes ON)

  • Disable phases you want to skip (checkboxes OFF)

Common configurations:

  • Quick test: Phase 1 only (disable everything else)

  • Fast iteration: Phase 1 + 2 only

  • Full quality: All phases enabled

Step 2: Configure Base Settings

Navigate to the SDXL Image Settings Controller node:

  • Set subdirectory for organized output (e.g., "4phase_test")

  • Choose resolution_preset (recommended: 1024×1024 for testing)

  • Set seed if desired.

  • Set seed_control to desired setting

  • Set batch_size to 1 for initial tests

Step 3: Load Your Models

Configure model loaders for each phase:

  • Phase 1 Model Loader: Select your SDXL-based checkpoint (SDXL, Pony, Illustrious, NoobAI, etc.)

  • Phase 2 Model Loader: Same or different SDXL-based model

  • Phase 3 Model Loader: Same or different SDXL-based model

  • Phase 4 Model Loader: Same or different SDXL-based model

Pro Tip: Using the same model across all phases yields maximum consistency. Using different models creates hybrid artistic effects.

Step 4: Configure Phase Settings

In the SDXL Image Settings Controller, the workflow has default settings that work for most models, but you may need to adjust them for specific checkpoints.

Denoise Settings (Most Important - Controls How Much Each Phase Transforms):

Phase 1: Recommended setting is 1.0 (full generation from noise)

Understanding Denoise Ranges:

  • 0.20-0.35: Minimal change, maximum detail preservation - subtle refinement of existing image

  • 0.35-0.55: Moderate refinement - noticeable improvements while maintaining composition

  • 0.55-0.70: Significant transformation - major changes while keeping general style

  • 0.70-1.0: Nearly complete to complete change - rebuilds the image but maintains style characteristics

For Complete Style Changes Between Models: When switching models to completely change the style (anime to realistic, or vice versa), you need denoise values of 0.45 to 0.60. Lower values may not give the model enough freedom to transform the style. Higher values might lose too much of the original composition.

You can set any phase to any denoise value depending on what you want to achieve. Higher denoise = more transformation. Lower denoise = more preservation of the previous phase.

Other Settings (Model-Dependent):

The workflow defaults should work for most models:

  • Steps, CFG, Sampler, Scheduler vary by checkpoint

  • Test with workflow defaults first

  • Adjust if results aren't as expected

  • Some models prefer specific samplers (check model card)

Step 5: Configure Prompts

Locate the Multi-Phase Prompt Manager (4 Phase) V2 node:

Phase 1:

  • Enter your base prompt in phase_1_positive_prompt

  • Enter negative prompt in phase_1_negative_prompt

Phase 2/3/4:

  • Each phase has a mode dropdown (default: "Append to Base")

  • Set mode: Use Base, Use Previous, Unique Prompt, Append to Base, or Append to Previous

  • Enter custom text in the positive/negative custom fields if using Unique or Append modes

Step 6: Configure LoRAs (Optional)

Load LoRAs:

  • Use LoRA Manager to visually select LoRAs for each phase

  • Each phase has its own Lora Loader node

  • LoRAs can be different per phase or the same across all phases

Select Trigger Words:

  • Each phase has TriggerWord Toggle nodes

  • Enable/disable specific trigger words for that phase

  • Triggers connect automatically to the Multi-Phase Prompt Manager

Carryover Trigger Words:

  • Phase 2, 3, and 4 have "Carryover TriggerWord Toggle" nodes

  • These let you select trigger words from PREVIOUS phases to carry forward

  • Example: Enable a Phase 1 trigger in Phase 2's carryover node to keep that effect

  • This gives you control over which LoRA effects persist across phases

Step 7: Run Your First Generation

  • Click "Queue Prompt" in ComfyUI

  • Monitor console for any errors

  • Check ShowText nodes for configuration verification

  • Find outputs in your configured subdirectory

Important: Embedded Workflow Information

ComfyUI automatically embeds the complete workflow, settings, prompts, models, seed, and LoRAs into every generated image's metadata.

Why Run Once Before Changing Settings:

  • Generate one image with the default workflow settings first

  • This image contains the original workflow configuration

  • If you make changes and want to return to defaults, simply drag this image back into the ComfyUI GUI

  • The workflow will reload with all original settings: prompts, models, seed, LoRAs, and all node configurations

  • This is your "reset to defaults" backup

Benefits:

  • Never lose working configurations

  • Easy to compare different settings

  • Share complete workflows by sharing images

  • Document exactly what created each image

Loading Workflow from Generated Images

Once you've generated images with this workflow:

  1. Find any image generated by this workflow in your output folder

  2. Drag the image directly onto the ComfyUI canvas

  3. The complete workflow loads with all settings from that generation

  4. All prompts, models, seed, LoRAs, and node settings are restored exactly as they were

This is the easiest way to reload the workflow or share configurations with others.

🎯 Creative Possibilities

Multi-Phase Transformation

Phase 1: Your Starting Point

  • Generate your base concept at any resolution

  • Set the foundation for what comes next

  • Use any SDXL checkpoint you want

  • Add LoRAs to establish your style

  • Save point: Your original vision

Phase 2: First Evolution

  • Keep refining the same concept, OR...

  • Take it in a completely different direction

  • Switch models for style hybridization

  • Add different LoRAs for new effects

  • Choose your denoise: 0.3 for tweaks, 0.6 for transformation

  • Save point: See how it evolved

Phase 3: Deep Dive

  • Continue the journey or pivot again

  • Optional FaceDetailer if you want perfect faces

  • Mix in more LoRAs—you've got independent control

  • This is where details really start to pop

  • Save points: Before and after face enhancement

Phase 4: The Final Form

  • Polish everything to perfection, or...

  • Make one last dramatic transformation

  • Second FaceDetailer pass for ultimate facial control

  • 4x upscaling option for massive detail

  • Optional third FaceDetailer on the upscaled version

  • Save points: Base, FaceDetailer, Upscaled, Upscale+FaceDetailer

Creative Control Systems

Play With Everything:

  • Each phase gets its own prompt—inherit from Phase 1 or go rogue

  • 12+ LoRAs across all phases? Go for it. No cross-contamination.

  • Mix photorealistic and anime models for hybrid styles

  • Randomize seeds and generate hundreds of variations

  • Save every phase to see the complete evolution

Resolution Freedom:

  • 30+ presets from 512×512 to 2688×1792

  • Square, portrait, landscape, ultra-wide

  • Start small for speed, go huge for detail

Seed Modes for Different Vibes:

  • Fixed: Perfect the same concept across phases

  • Randomize: Chaos mode—every generation is unique

  • Increment/Decrement: Subtle variations on a theme

Batch Generation:

  • Generate 1-50+ images at once

  • Randomize everything for variety

  • Use wildcards for prompt variations

  • See what emerges from creative chaos

💡 Creative Modes

Strategic LoRA Usage Across Phases

You can apply different LoRAs and trigger words in each phase for precise control over the generation process:

Phase-Specific LoRA Strategy:

  • Phase 1: Shape/structure LoRAs - body types, armor, clothing, pose

  • Phase 2: Detail enhancement LoRAs - Wonderful Details, texture enhancers, shapers (Hands, items, etc.).

  • Phase 3: Lighting and atmosphere LoRAs - dramatic lighting, effects, ambiance, liquid, smoke light shapers, and enhancers.

  • Phase 4: Refiner LoRAs - final polish, quality boosters, detailers, and refiners.

Why This Works:

  • Build the foundation first (structure and form)

  • Add complexity progressively (details and textures)

  • Apply finishing touches last (lighting and refinement)

  • Each phase focuses on one aspect without overwhelming earlier stages

Using Carryover Triggers:

  • Use carryover nodes to keep essential triggers from previous phases

  • Drop triggers that served their purpose: Example: Keep body type trigger through all phases, but drop clothing detail trigger after Phase 2

  • Enable LoRAs used before with weights to add heavy weight to those LoRAs.

Strategic Prompt Adjustments Across Phases:

Your Phase 1 prompt establishes the complete scene - subject, background, atmosphere, everything. Later phases refine and emphasize what's already there.

What You CAN Do in Later Phases:

  • Adjust weights on specific elements:

    • Emphasize colors: "(red umbrella:1.4)" to intensify existing red

    • Boost specific details: "(intricate armor:1.3)", "(glowing moon:1.5)"

    • De-emphasize elements: "(forest background:0.8)" to push it back

  • Append to base prompt:

    • Add emphasis terms: "dramatic lighting", "highly detailed armor"

    • Quality adjustments: "8k", "masterpiece", "professional photography"

    • Atmosphere enhancements: "volumetric fog", "rim lighting"

  • Slightly modify the prompt:

    • Shift focus while keeping the scene: adjust what gets attention

    • Change quality descriptors for different models

    • Adjust atmosphere while keeping composition

  • Change models to shift style while maintaining composition

  • Apply different LoRAs to enhance specific aspects (details, lighting, quality)

  • Use different embeddings per model or for quality adjustments

The Key Rule: Stay true to what Phase 1 established. You're refining, emphasizing, and enhancing the existing scene - not adding new elements that weren't there from the start.

Example:

  • Phase 1: "dark elf warrior, red umbrella, ornate black kimono, blonde braid, forest background, night, glowing moon, falling petals"

  • Phase 2: Same base + "(intricate kimono patterns:1.3)" - emphasizes existing detail

  • Phase 3: Same base + "(dramatic red lighting:1.4), volumetric atmosphere" - enhances existing lighting

  • Phase 4: Same base + "Amateur Photography, harsh lighting, High contrast lighting, 80s Poloroid Photo”, quality change or enhancement.

Maximum Detail Mode

When to use:

  • You've got a prompt you absolutely love

  • You want to extract every possible detail

  • Portfolio-worthy final renders

  • That one perfect character/scene

How to set it up:

  • Enable all four phases

  • Use the same model throughout (consistency)

  • Progressive denoise: 1.0 → 0.35 → 0.35→ 0.25

  • Enable all FaceDetailer passes, you can change detail different parts in each phase as well. Defaults are face and skin or face and person in Phase 4

  • Enable upscaling for 4K

What you get: 12 saved images showing the complete evolution from rough concept to masterpiece

Style Transformation Mode (This is where it gets FUN)

When to use:

  • You want to see your concept in different styles

  • Experimenting with model hybrids

  • Creating something totally unique

  • "What if this photorealistic portrait was anime?"

How to set it up:

  • Phase 1: Start with any model

  • Phase 2: Switch to a completely different model

  • Phase 3: Try another model or stick with Phase 2

  • Phase 4: Go wild

  • Use "Unique Prompt" mode for each phase

  • Higher denoise values (0.40-0.60) for bigger transformations

  • Different LoRA sets per phase

Example journeys:

  • Phase 1: Photorealistic → Phase 2: Pony Diffusion → Phase 3-4: Full anime

  • Phase 1: Anime base → Phase 2-4: Progressive photorealism

  • Phase 1: SDXL base → Phase 2: Illustrious → Phase 3: NoobAI → Phase 4: Back to SDXL

Anime to Realistic Transformation Strategy:

This is a powerful technique for creating unique hybrid styles:

Phase 1 (Anime Model - Illustrious, NoobAI, Pony):

  • Denoise: 1.0

  • LoRAs: Body type, armor, clothing, pose (structure)

  • Build the composition and character in anime style

Phase 2 (Switch to Semi-Realistic or Realistic Model):

  • Denoise: 0.55 (significant transformation while keeping composition)

  • Model: SDXL realistic checkpoint or semi-realistic blend

  • LoRAs: Detail enhancers (Wonderful Details, texture enhancers)

  • This is where the style shift happens

Phase 3 (Continue with Realistic Model):

  • Denoise: 0.35-0.45

  • LoRAs: Lighting and atmosphere LoRAs

  • Refine the realistic elements, add depth

Phase 4 (Final Realistic Polish):

  • Denoise: 0.25-0.35

  • LoRAs: Refiner LoRAs, quality boosters

  • Final photorealistic details

Why This Works:

  • Anime models excel at composition, character design, and creative poses

  • Realistic models excel at texture, lighting, and physical accuracy

  • The 0.55 denoise in Phase 2 maintains the composition while transforming the style

  • Progressive LoRA strategy builds complexity without overwhelming the transformation

  • Result: Compositions that feel anime-inspired but look photorealistic

What you get: Mind-bending style evolution that you can't achieve any other way

Progressive Enhancement Mode

When to use:

  • You like your Phase 1 but want MORE

  • Refining details without changing the core concept

  • Testing how far refinement can go

  • Learning what denoise values do

How to set it up:

  • Enable Phases 1 and 2, bypass 3 and 4 at first

  • Use "Use Base (Phase 1)" mode for consistency

  • Try different Phase 2 denoise values to see effects

  • Same model both phases

  • Optional: Enable Phase 3 to see deeper refinement

What you get: Perfect for learning how the workflow behaves and finding your sweet spot

Chaos Generation Mode

When to use:

  • You want to discover happy accidents

  • Building a dataset of diverse images

  • Looking for inspiration

  • "Let's see what happens"

How to set it up:

  • Seed: "randomize"

  • Use wildcards in prompts for variety

  • Maybe randomize denoise slightly per phase

  • Mix and match different models/LoRAs

  • Enable everything

What you get: Hundreds of unique images, many you never would have thought to create manually. This is where you find your best work.

LoRA Experimentation Mode

When to use:

  • You've got a pile of new LoRAs to test

  • Want to see how LoRAs interact

  • Building your "LoRA toolkit" understanding

  • Creating signature multi-LoRA combinations

How to set it up:

  • Phase 1: Test 3-4 foundational LoRAs

  • Phase 2: Add different LoRAs to see how they mix

  • Phase 3: Try yet another combination

  • Phase 4: Your favorite combo from earlier phases

  • Use ShowText nodes to verify which triggers are active

  • Take notes on what works together

What you get: Deep understanding of how to orchestrate multiple LoRAs for unique styles

The "Perfect One Image" Mode

When to use:

  • Commission work or gift art

  • That character reference you need to nail

  • Your OC that has to be exactly right

  • Profile picture / avatar that represents you

How to set it up:

  • Use a fixed seed you've tested

  • All phases enabled with carefully chosen models

  • Hand-select LoRAs that nail the aesthetic

  • Multiple FaceDetailer passes

  • Full upscale to 4K

  • Take your time adjusting prompts between test runs

What you get: The single most refined image possible from this workflow

🖼️ Real Example: 4-Phase Multi-Model Transformation

Here's an actual generation showing what this workflow can do when you mix different models across phases (see example posts):

The Setup:

  • Started with a dark elf warrior concept

  • Used 3 different models across the 4 phases

  • Progressive denoise pattern for controlled evolution

  • Each phase refined details while maintaining core composition

What Happened Across Phases:

Phase 1 (Base Generation): Established the foundation—character pose, composition, dramatic red lighting, the umbrella, the flowing fabric. This is where the core concept lives.

Phase 2 (First Model Switch): Notice how the facial details started to sharpen, the fabric textures became more defined, and the ornate patterns on the clothing gained clarity. The red light effects intensified while maintaining the original pose perfectly.

Phase 3 (Continued Refinement): The intricate embroidery on the kimono got more detailed, the skin texture refined, individual strands of the braid became visible, and the background elements (falling petals, atmospheric particles) gained definition. The golden accents and ornamental details really popped.

Phase 4 (Final Polish): Ultimate refinement—every detail crisp, the contrast between the dark skin and golden ornaments perfected, the red glow effects optimized, fabric folds and patterns at maximum detail. This is the keeper.

The Magic: Same seed, same basic prompt structure, but THREE different models shaped this image across four phases. The result? A style hybrid that doesn't exist in any single checkpoint. The photorealistic rendering combined with painterly lighting and anime-inspired composition elements—this only happens with multi-phase model mixing.

Key Takeaway: Notice how each phase REFINED rather than REPLACED. The composition from Phase 1 stayed rock-solid through all four phases. That's the beauty of controlled denoise values—you get evolution, not revolution.

This is one image. Imagine doing this with batch_size 20 and randomized seeds. That's when things get really interesting.

🐛 Troubleshooting

Workflow Won't Load

Symptom: Error messages when loading workflow JSON

Solutions:

  1. Verify all custom node packs are installed

  2. Install the two required custom Python files

  3. Restart ComfyUI completely

  4. Check console for missing node errors

  5. Update ComfyUI to latest version

  6. Reinstall missing node packs via Manager

Custom Nodes Not Found

Symptom: "SDXL Image Settings Controller" or "Multi-Phase Prompt Manager" nodes missing

Solutions:

  1. Verify files are in custom_nodes/ folder (not in subfolder)

  2. Check file names match exactly:

    • sdxl_image_settings_controller.py

    • multi_phase_prompt_manager.py

  3. Open files in text editor, verify no syntax errors

  4. Restart ComfyUI after copying files

  5. Check Python console for import errors

Out of VRAM

Symptom: Generation crashes mid-process, CUDA out of memory errors

Solutions:

  1. Reduce batch_size to 1

  2. Bypass Phase 3 and Phase 4 temporarily

  3. Disable all FaceDetailer passes

  4. Disable upscaling

  5. Use lower resolution preset

  6. Verify model unload nodes aren't bypassed

  7. Close other GPU-intensive applications

  8. Update to latest GPU drivers

Images Not Saving

Symptom: Workflow runs but no output images appear

Solutions:

  1. Check subdirectory path in controller

  2. Verify SaveImage nodes aren't bypassed

  3. Check ComfyUI output folder permissions

  4. Review console for file write errors

  5. Ensure drive has sufficient space

  6. Try removing subdirectory (use default output)

Phase 2/3/4 Look Wrong

Symptom: Later phases don't match expectations

Solutions:

  1. Check prompt mode in Multi-Phase Prompt Manager

  2. CFG Settings are too high for the selected model.

  3. Verify denoise values (too high = too much change)

  4. Review ShowText nodes for actual prompts used

  5. Confirm correct model loaded for each phase

  6. Check LoRA weight values and triggers are selected correctly

  7. Verify VAE consistency across phases

  8. Try "Use Base (Phase 1)" mode for consistency

LoRA Effects Not Working

Symptom: LoRA triggers not producing expected results

Solutions:

  1. Verify trigger words match LoRA requirements

  2. Check trigger selection in prompt manager

  3. Confirm LoRA Manager has LoRAs loaded

  4. Review ShowText outputs showing triggers

  5. Test LoRAs individually in simple workflow

  6. Verify LoRA weight values (typically 0.6-1.0)

  7. Check for trigger typos

Workflow Runs Slowly

Symptom: Generation takes much longer than expected

Solutions:

  1. To go through all 4 Phases expect 2 minutes or more.

  2. Face detailers will detect and detail EVERY face in the image and that will slow generation down tremendously.

  3. Disable optional features for faster iteration:

    • Bypass Phase 3 and Phase 4 for testing

    • Disable FaceDetailer passes

    • Disable upscaling

  4. Reduce steps per phase (20/15/12/12). Though, this will affect quality of images.

  5. Use lower resolution for prompt testing

  6. Test on faster scheduler (euler vs dpmpp)

  7. Ensure SSD storage (not HDD)

FaceDetailer Not Detecting Faces

Symptom: FaceDetailer nodes run but faces unchanged

Solutions:

  1. Verify face is clearly visible and forward-facing

  2. Check bbox detection model loaded correctly

  3. Adjust FaceDetailer bbox_threshold (try 0.5)

  4. Verify SAM model loaded properly

  5. Check detection_hint is set to "enabled"

  6. Try different face at different angle

  7. Ensure face is reasonably sized in composition

The 80/20 Rule

  • 80% of your best images come from 20% of your generations

  • That's NORMAL

  • Generate more

  • Sort ruthlessly

  • Only show the masterpieces

Don't get discouraged by the 80%. They're part of the process.

📚 Additional Resources

Documentation

Included Files:

  • SDXL_4_Phase_Super_Refiner_V1.json - Main workflow

  • sdxl_image_settings_controller.py - Settings controller node

  • multi_phase_prompt_manager.py - Prompt manager node

  • vae_selector.py

  • This README

📄 License & Credits

Workflow Designer: DarthRidonkulous
Custom Nodes: Developed with AI assistance (Claude, Anthropic)
Testing: Extensive production validation across model families. Thousands of images generated in testing.

Usage Terms

✅ Allowed:

  • Free to use and modify for personal projects

  • Free to use for commercial projects

  • Share and distribute modified versions

  • Use custom nodes in other workflows

  • Create derivative workflows

❌ Not Allowed:

  • Claiming original authorship

  • Selling workflow as-is without modification

  • Removing attribution from custom nodes

Attribution

Credit appreciated but not required. If sharing your results:

  • Mention "4-Phase Super Refiner by DarthRidonkulous"

  • Link to CivitAI post

  • Tag @DarthRidonkulous on social media

  • Share in CivitAI comments section

💬 Support the Creator

If you find this workflow valuable:

Share Your Work:

  • Post results on CivitAI with workflow credit

  • Share on social media with attribution

  • Help other users in comments section

Provide Feedback:

  • Leave detailed reviews

  • Report bugs constructively

  • Suggest features for future versions

  • Share your configuration discoveries

Join the Community:

  • Follow on CivitAI for updates

  • Participate in discussions

  • Share workflow tips and tricks

  • Help troubleshoot for other users

🙏 Acknowledgments

Special Thanks:

  • ComfyUI development team for the platform

  • Custom node developers for extensibility

  • Beta testers for feedback and validation

  • CivitAI community for support and inspiration

  • Model creators for excellent SDXL checkpoints

Built With:

  • ComfyUI workflow framework

  • Impact Pack for FaceDetailer

  • LoRA Manager for visual interface

  • rgthree nodes for Power Prompts

  • pysssss scripts for utilities

🎉 Let's Create Something Amazing

The SDXL 4-Phase Super Refiner V1 is about one thing: giving you creative superpowers.

This isn't a corporate production tool. This is a playground where you can:

  • Mix styles that shouldn't exist

  • Transform reality into fantasy (or vice versa)

  • Stack 12 LoRAs and see what happens

  • Generate hundreds of variations on a theme

  • Discover your signature aesthetic

  • Create art that's uniquely YOURS

This Workflow is For:

The Experimenters who want to push boundaries and see what's possible

The Artists who have a vision and need the tools to realize it

The Collectors who want to generate massive galleries of diverse art

The Perfectionists who want every detail exactly right

The Chaos Agents who randomize everything and sort through gold later

The Style Hunters looking to create hybrid aesthetics no one else has

The Character Designers who need consistency across multiple refinements

The NSFW Creators who need privacy and control (we see you, no judgment)

What Makes This Different

Most workflows give you ONE shot to get it right. This one gives you FOUR chances to:

  • Refine your vision

  • Change direction completely

  • Experiment with different models

  • Layer different LoRA combinations

  • Build complexity progressively

  • Save every phase of evolution

It's not about following a formula. It's about having the freedom to create however YOU want.

The Community Aspect

Share your discoveries:

  • Post your wildest style combinations

  • Show your before/after transformations

  • Explain your LoRA orchestration

  • Share your favorite denoise patterns

  • Help others learn from your experiments

Learn from others:

  • Check the comments for creative techniques

  • See how people are using different model combos

  • Find new LoRAs to experiment with

  • Discover approaches you hadn't considered

Contribute back:

  • Share workflow modifications

  • Post your settings for specific styles

  • Help troubleshoot issues

  • Celebrate each other's creations

A Note About "Production"

Yes, this workflow CAN produce professional-grade, print-ready, portfolio-quality images. But that's not WHY it exists.

It exists because creativity thrives with:

  • Freedom to experiment

  • Control over every variable

  • Flexibility to change direction

  • Power to realize any vision

  • Tools that get out of your way

If you happen to sell your art, great. If you're making references for your game, awesome. If you're just creating for the pure joy of it, that's perfect too.

What You Should Do Next:

  • Download the workflow

  • Start with Phase 1 only

  • Get comfortable, then get weird

  • Share what you create

  • Help others in comments

  • Push the boundaries and see what breaks

Final Thoughts

Hundreds of hours went into building this workflow, testing it, breaking it, fixing it, and refining it. Not to create a production tool, but to create a creative instrument.

Think of it like a guitar—it has strings, frets, and pickups. But what you PLAY on it? That's all you.

This workflow has phases, denoise values, and LoRA controls. What you CREATE with it? That's all you.

So go create something weird. Something beautiful. Something that shouldn't exist but does because you made it exist.

That's what this is all about.

See you in the comments. Show me what you make. 🚀

Workflow Version: 1.0
Documentation Version: 1.0
Last Updated: January 2025
Created by: DarthRidonkulous
Built with: ComfyUI + hundreds of hours of experimentation
Made for: Creators, experimenters, and artists who want MORE

此模型生成的图像

未找到图像。