ComfyUI SDXL Guide: Setup, Refiner Workflow & Best Practices

What is SDXL?

Stable Diffusion XL (SDXL) is Stability AI's high-resolution image generation model. Compared to SD 1.5, SDXL produces significantly better image quality, handles complex compositions well, and generates more coherent text in images.

Key differences from SD 1.5:

Setup

Model Download

Place both in ComfyUI/models/checkpoints/.

Resolution

SDXL was trained at 1024x1024. Always use resolutions that total approximately 1 megapixel:

Basic Text-to-Image Workflow

The simplest SDXL workflow is identical to SD 1.5:

1. Load Checkpoint → SDXL Base model
2. CLIP Text Encode (positive) → your prompt
3. CLIP Text Encode (negative) → elements to avoid
4. Empty Latent Image → 1024x1024
5. KSampler → steps: 25–30, cfg: 6–8
6. VAE Decode → Save Image

Base + Refiner Workflow

SDXL's unique feature is the two-stage generation pipeline. The Base model handles the main composition and structure, then the Refiner model improves fine detail and texture quality in a second pass.

How It Works

1. The Base model generates for the first portion of sampling steps (e.g., 20 out of 25 total)
2. The Refiner takes over for the remaining steps, refining details without changing composition

Node Setup

1. Load Checkpoint (Base) → SDXL Base
2. Load Checkpoint (Refiner) → SDXL Refiner
3. CLIP Text Encode (x2) → connected to Base's CLIP
4. Empty Latent Image → 1024x1024
5. KSampler (Base) → steps: 25, end_at_step: 20
6. KSampler (Refiner) → steps: 25, start_at_step: 20, connects to Base's latent output
7. VAE Decode → Save Image (using Refiner's VAE)

When to Skip the Refiner

The Refiner adds detail but also adds processing time and VRAM usage. Skip it when:

• You're doing quick iterations or testing prompts
• Your GPU has limited VRAM
• The base output already looks good

Many community checkpoints fine-tuned from SDXL don't benefit from the refiner.

SDXL ReVision: Image-Guided Generation

ReVision lets you use images as conceptual input instead of (or alongside) text prompts. It extracts visual concepts from reference images and generates new images inspired by them.

Setup

Download the CLIP-G Vision model:

• clip_vision_g.safetensors → place in ComfyUI/models/clip_vision/

Workflow

1. Load Image → your reference image(s)
2. CLIP Vision Encode → extracts visual features
3. unCLIP Conditioning → applies features as conditioning
4. KSampler → generates inspired by the reference

You can combine multiple reference images by chaining unCLIP Conditioning nodes. The strength parameter controls how strongly each image influences the output.

Common Issues and Fixes

Output quality is poor or blurry

• Verify you're generating at ~1MP resolution (1024x1024 or equivalent)
• 512x512 will produce very poor results with SDXL

VRAM too low for Base + Refiner

• Use the Base model alone — results are still good
• Or use a community fine-tuned SDXL checkpoint that doesn't need a refiner

LoRA not working with SDXL

• Make sure the LoRA is specifically made for SDXL — SD1.5 LoRAs are incompatible
• SDXL LoRAs are labeled accordingly on Civitai and HuggingFace

Colors look washed out

• Try a different VAE — some SDXL checkpoints benefit from an external VAE
• The SDXL Refiner's VAE often produces better colors

Related Guides

• Text to Image — Basic generation concepts
• Flux Guide — Next-generation model with even better quality
• SD 3.5 Guide — Latest Stability AI model
• LoRA Guide — Fine-tuning with adapters