<h1 id="image_z_image_sdxl_refiner">image_z_image_SDXL_Refiner</h1>This ComfyUI workflow generates or processes images using ZImage Base model (Phase 1), refines them with SDXL-based models (Phase 2), applies automatic face and skin enhancement, followed by dual-stage upscaling.<h2 id="installation">Installation</h2><h3 id="required-custom-nodes">Required Custom Nodes</h3>Install via ComfyUI Manager:<ol><li>ComfyUI-Easy-Use - https://github.com/yolain/ComfyUI-Easy-Use</li><li>ComfyUI-Levelpixel - https://github.com/levelpixel/ComfyUI-Levelpixel</li><li>ComfyUI-Impact-Pack - https://github.com/ltdrdata/ComfyUI-Impact-Pack</li><li>ComfyUI-Impact-Subpack - https://github.com/ltdrdata/ComfyUI-Impact-Subpack</li><li>ComfyUI KJ Nodes - https://github.com/kijai/ComfyUI-KJNodes</li><li>ComfyUI LoRA Manager - https://github.com/cubiq/ComfyUI_Lora_Manager</li><li>ComfyUI-pysssss - https://github.com/pythongosssss/ComfyUI-Custom-Scripts</li><li>ComfyUI Comfyroll Custom Nodes - https://github.com/Suzie1/ComfyUI_Comfyroll_CustomNodes</li></ol><h3 id="required-model-files">Required Model Files</h3>Detection Models (ComfyUI/models/ultralytics/):<ul><li><code>bbox/face_yolov8n_v2.pt</code> - Face detection</li><li><code>segm/skin_yolov8n-seg_800.pt</code> - Skin segmentation</li></ul>SAM Model (ComfyUI/models/sams/):<ul><li><code>sam_vit_b_01ec64.pth</code> - Segment Anything Model</li></ul>Checkpoints (ComfyUI/models/checkpoints/ or ComfyUI/models/unet/):<ul><li>ZImage Base checkpoint (for Phase 1)</li><li>SDXL Refiner or any SDXL-based checkpoint (for Phase 2)</li></ul>Upscale Models (ComfyUI/models/upscale_models/):<ul><li>RealESRGAN_x4plus.pth or similar</li></ul>Detection models and SAM download automatically on first use or from: https://github.com/ltdrdata/ComfyUI-Impact-Pack<h2 id="workflow-structure">Workflow Structure</h2><h3 id="input-selection">Input Selection</h3>Uses CR Latent Input Switch to choose between:<ul><li>Input 1: Uploaded image (LoadImage → VAEEncode)</li><li>Input 2: Empty latent for generation from scratch (default: 896x1152)</li></ul><h3 id="phase-1:-zimage-base-generation">Phase 1: ZImage Base Generation</h3><ul><li>Processes selected input using ZImage Base checkpoint</li><li>Default settings: 50 steps, CFG 5, uni_pc_bh2 sampler, ddim_uniform scheduler</li><li>Supports LoRAs via first Lora Loader</li><li>Output goes to Phase 2</li></ul><h3 id="phase-2:-sdxl-refiner">Phase 2: SDXL Refiner</h3><ul><li>Refines Phase 1 output using SDXL-based checkpoint</li><li>KSamplerAdvanced settings: 50 steps, CFG 1.9, start step 40</li><li>Sampler: dpmpp_3m_sde_gpu, Scheduler: beta57</li><li>Supports multiple LoRAs via "Phase 2 Lora Loader"</li><li>Trigger words managed by TriggerWord Toggle node</li></ul><h3 id="detailing-system">Detailing System</h3><ul><li>Automatic face detection: YOLOv8n v2 (bbox/face_yolov8n_v2.pt)</li><li>Skin segmentation: YOLOv8n-seg (segm/skin_yolov8n-seg_800.pt)</li><li>SAM model for precise mask generation</li><li>FaceDetailer settings: 25 steps, CFG 6, denoise 0.25, bbox_threshold 0.3</li></ul><h3 id="upscaling">Upscaling</h3><ul><li>Two-stage progressive upscaling</li><li>Uses RealESRGAN or similar models</li><li>Each stage independently controlled via Fast Groups Bypasser</li></ul><h2 id="model-compatibility">Model Compatibility</h2>Phase 1: Requires ZImage Base checkpointPhase 2: Accepts any SDXL-architecture checkpoint:<ul><li>Official SDXL Refiner</li><li>SDXL base checkpoints</li><li>Pony-based models</li><li>Illustrious-based models</li><li>Other SDXL derivatives</li></ul>Important: Different SDXL variants may require different sampler/scheduler settings. The workflow uses dpmpp_3m_sde_gpu with beta57 scheduler for Phase 2, and uni_pc_bh2 with ddim_uniform for Phase 1. For Pony or Illustrious models, you may need to adjust:<ul><li>Scheduler (try karras, normal, simple)</li><li>Sampler (try euler_a, dpmpp_2m)</li><li>CFG scale and step counts</li></ul><h2 id="usage">Usage</h2><h3 id="basic-setup">Basic Setup</h3><ol><li>Set output folder: Enter name in "Save Subdirectory Name"</li><li>Choose input: CR Latent Input Switch - 1 for uploaded image, 2 for generation</li><li>Load models: ZImage Base for Phase 1, SDXL model for Phase 2</li><li>Set prompts: Phase 1 prompts for generation, Phase 2 prompts for refinement</li><li>Configure LoRAs: Load in respective Lora Loader nodes, toggle trigger words</li></ol><h3 id="fast-groups-bypasser">Fast Groups Bypasser</h3>Control workflow sections:<ul><li>Phase 1 - ZImage Base Generation: Main generation (keep enabled)</li><li>Phase 2 - SDXL Refiner: Refinement pass (keep enabled)</li><li>Model Unload, Clear Cache and VRAM: Enable if low VRAM (default: disabled)</li><li>Detailer Bridge: Prepares for face/skin enhancement (keep enabled)</li><li>Upscale 1: First upscale pass (disable to skip)</li><li>Upscale 2: Second upscale pass (disable to skip)</li></ul><h3 id="output-files">Output Files</h3>All images save to: ComfyUI/output/[subdirectory]/Includes complete metadata: prompts, seeds, steps, CFG, models, LoRAs, all workflow settings.<h2 id="default-settings">Default Settings</h2>Phase 1 (ZImage Base):<ul><li>Steps: 50</li><li>CFG: 5</li><li>Sampler: uni_pc_bh2</li><li>Scheduler: ddim_uniform</li></ul>Phase 2 (SDXL Refiner):<ul><li>Steps: 50</li><li>CFG: 1.9</li><li>Start step: 40</li><li>Sampler: dpmpp_3m_sde_gpu</li><li>Scheduler: beta57</li></ul>FaceDetailer:<ul><li>Steps: 25</li><li>CFG: 6</li><li>Denoise: 0.25</li><li>bbox_threshold: 0.3</li></ul>Empty Latent: 896 x 1152<h2 id="troubleshooting">Troubleshooting</h2>Out of VRAM errors:<ul><li>Enable Model Unload/Clear Cache group via Fast Groups Bypasser</li><li>Disable one or both upscale stages</li><li>Lower step counts</li></ul>Face detailer not activating:<ul><li>Lower bbox_threshold (default is 0.3, try 0.2)</li><li>Ensure faces are clearly visible and adequately sized</li><li>Verify detector model files downloaded correctly</li></ul>Using non-standard SDXL models (Pony, Illustrious):<ul><li>Adjust Phase 2 sampler/scheduler settings</li><li>Common alternatives: euler_a sampler with karras scheduler</li><li>Test different CFG values</li><li>Check model card for recommended settings</li></ul><h2 id="technical-details">Technical Details</h2>Optimized for: RTX 4090 with 24GB VRAM Processing flow:<ol><li>Select input (uploaded image or empty latent)</li><li>Phase 1: ZImage Base generation/processing</li><li>Phase 2: SDXL refinement</li><li>Face and skin region detection</li><li>Targeted detail enhancement</li><li>Progressive dual upscaling</li><li>Save with complete metadata</li></ol>Execution time: 40 seconds to 3 minutes depending on hardware, settings, and enabled stages. 40 seconds with current settings on a RTX 4090 GPU.

Files with SHA256: 6b869d86da6c5e7de33e697066cf7512941895f07cae4ad8990cc05d80a73dd4