Looking to enhance your AI image generation capabilities? SDXL (Stable Diffusion XL) represents a significant leap forward in text-to-image models, offering improved quality and capabilities compared to earlier versions. This guide will walk you through installing and running SDXL models in ComfyUI, a powerful open-source interface for AI image generation. It also answers the question most readers arrive with: which SDXL checkpoints are actually worth loading in ComfyUI in 2026.
If you have not installed ComfyUI itself yet, start with our ComfyUI installation guide and come back here once the interface opens in your browser.
(photorealistic), beautiful lighting, best quality, realistic, full body portrait, real picture, intricate details, depth of field, 1girl, in a cold snowstorm, wearing winter camo military fatigues, camo plate carrier rig, combat gloves, (magazine pouches), (kneepads), highly-detailed, perfect face, blue eyes, lips, wide hips, small waist, tall, makeup, tactical, Fujifilm XT3, outdoors, bright day, Beautiful lighting, RAW photo, 8k uhd, film grain, ((bokeh))
Steps: 20, Sampler: DPM++ 2M Karras, CFG scale: 7
Seed: 3594172103
Prerequisites
SDXL at its native 1024x1024 resolution runs comfortably on 8 GB of VRAM and gets faster and more flexible from 12 GB upward. Before we begin, make sure you have:
- A computer with at least 8GB VRAM (12GB+ recommended)
- ComfyUI installed and running
- Basic familiarity with downloading and managing model files
- Roughly 7 GB of free disk space per SDXL checkpoint, plus a little more for the VAE and any LoRAs
Understanding SDXL Model Types
SDXL comes in several variants:
- Base SDXL 1.0: The standard model offering excellent image quality
- SDXL Turbo: Optimized for speed with slightly lower quality
- SDXL Lightning: A balanced option between speed and quality
The base model is what community fine-tunes are built on, and those fine-tunes are what most people actually run day to day. Turbo and Lightning variants trade a few steps for speed and want much lower step counts and CFG values than the base model, so always follow the settings on the model card when you use them.
Eg. example of the variants:
Best SDXL Checkpoints for ComfyUI in 2026
The short answer: load Juggernaut XL for photorealism, RealVisXL for clean product and portrait renders, AAM XL AnimeMix for anime, and DreamShaper XL when you want a single checkpoint that handles everything. All four are SDXL fine-tunes, so every workflow in this guide applies to them without changes.
| Model | Best for | Notes |
|---|---|---|
| Juggernaut XL | Photorealism | The default realistic checkpoint; strong skin, lighting and anatomy |
| RealVisXL | Realistic people and objects | Often edges Juggernaut on clean product and commercial shots |
| AAM XL AnimeMix | Modern anime | The go-to anime SDXL checkpoint |
| DreamShaper XL | All-purpose / semi-real | Flexes across photoreal, illustration and anime in one model |
Download the .safetensors file from Civitai or Hugging Face, read the model card for its recommended sampler, step count and CFG, and drop the file into models/checkpoints. For a deeper breakdown of each family, including anime alternatives and how to choose by VRAM, see our guide to the best SDXL models in 2026. If you want to compare SDXL against the newer Flux family before committing, read SDXL vs Flux in 2026.
Installation Steps
1. Download Required Files
First, download these essential components:
- SDXL base model checkpoint (or one of the fine-tunes above)
- SDXL VAE (Variational Autoencoder)
- CLIP text encoders
In practice, a standard SDXL checkpoint already bundles its VAE and both CLIP text encoders, so a single checkpoint file is enough to generate. A separate VAE file is only needed when a model card says so, or when you want to override the baked-in VAE (see the VAE notes below).
2. File Organization
Place the downloaded files in their respective ComfyUI directories:
models/checkpoints/ # For base model and fine-tunes
models/vae/ # For a standalone VAE file
models/clip/ # For CLIP encoders (only if shipped separately)
models/loras/ # For LoRA files
Putting a .safetensors file in the wrong folder is the single most common reason the Load Checkpoint node shows nothing. If a model does not appear in a dropdown, check the folder first, then press the refresh button in the ComfyUI interface.
3. ComfyUI Setup
Follow these steps to configure ComfyUI:
- Launch ComfyUI
- Update to the latest version
- Verify model detection by opening a Load Checkpoint node and confirming your file is listed
Creating Your First SDXL Workflow
The default text-to-image graph is all you need for a first SDXL image. Here's a basic workflow to get started:
- Add a KSampler node
- Connect SDXL checkpoint loader
- Set up your prompt
- Configure generation parameters
The graph, node by node: Load Checkpoint outputs the model, CLIP and VAE. Two CLIP Text Encode nodes take your positive and negative prompts. Empty Latent Image sets the canvas size. KSampler runs the diffusion steps. VAE Decode turns the latent into pixels, and Save Image writes the file. Wire positive and negative prompts into KSampler, connect the latent, then chain KSampler into VAE Decode and Save Image, and press Queue Prompt.
Recommended settings for 1024x1024
SDXL was trained around a one-megapixel canvas, so start at 1024x1024 or an equivalent aspect ratio such as 1152x896 or 896x1152. Going much smaller produces soft, incoherent images, and going much larger without an upscaler invites duplicated subjects.
| Setting | Starting point | Notes |
|---|---|---|
| Resolution | 1024x1024 | Keep total pixels near one megapixel |
| Steps | 20 to 30 | Lightning and Turbo variants want far fewer |
| CFG | 5 to 8 | Fine-tunes usually like the lower half of this range |
| Sampler | DPM++ 2M Karras or Euler a | Check the model card for its preferred sampler |
| Batch size | 1 to start | Increase only when VRAM allows |
The base plus refiner workflow
SDXL was designed as a two-stage system: the base model lays down composition and the refiner adds fine detail in the last steps. In ComfyUI you build this with two Load Checkpoint nodes and two KSampler Advanced nodes. The first sampler runs the base model for most of the steps and passes its latent, still noisy, to the second sampler, which finishes the remaining steps with the refiner. Modern fine-tunes such as Juggernaut XL and DreamShaper XL are trained to produce finished images on their own, so most people skip the refiner entirely. Try it only if you are using the plain base model and want extra detail in skin, fabric or foliage.
VAE notes
The VAE decodes the latent into your final image, and SDXL's original VAE can produce washed-out colors or NaN errors when run in half precision on some GPUs. If your outputs come out desaturated or blank, load a fixed fp16 SDXL VAE into models/vae, add a Load VAE node and connect it to VAE Decode in place of the checkpoint's built-in VAE. Most 2026 fine-tunes already bake in a corrected VAE, which is why the standalone file is optional.
Optimizing Performance
Tips for better results:
- Adjust batch sizes based on VRAM
- Experiment with sampling methods
- Fine-tune CFG values
Memory tips for 8 GB cards
ComfyUI manages VRAM well on its own, but three switches help on smaller GPUs. Launch with --lowvram to offload parts of the model to system RAM at the cost of speed. Load the checkpoint in a lower precision such as fp8 when your card supports it, which cuts model memory roughly in half with a small quality cost. Keep batch size at 1 and generate at native resolution, then use a dedicated upscaler rather than a bigger canvas. Our guide to upscaling in ComfyUI covers the ESRGAN and Ultimate SD Upscale approaches.
Troubleshooting Common Issues
-
Model loading errors: the checkpoint is not listed, or Load Checkpoint fails. Confirm the file is in
models/checkpoints, has a.safetensorsor.ckptextension, finished downloading (compare the file size with the model page) and refresh the browser tab. - Red nodes when loading a workflow: a shared workflow references a custom node you have not installed. Install the missing node pack through ComfyUI Manager, restart ComfyUI and reload the workflow.
-
Black or blank images: almost always a VAE precision problem. Switch to a fixed fp16 SDXL VAE as described above, or try the
--fp32-vaelaunch flag. -
VRAM limitations: out-of-memory errors mid-generation. Lower the resolution to 1024x1024, set batch size to 1, close other GPU applications and add
--lowvram. - Generation speed concerns: SDXL needs more compute than SD 1.5. Use a Lightning or Turbo checkpoint for drafts, reduce steps, and keep the refiner off unless you need it.
- Soft or duplicated subjects: resolution too low or too high for the model. Return to a one-megapixel canvas.
Advanced Techniques
Once comfortable with basics, explore:
- LoRA integration: add a Load LoRA node between Load Checkpoint and the CLIP Text Encode nodes, and see our guide to LoRAs in ComfyUI for stacking and troubleshooting
- Custom workflows: ControlNet, IP Adapter and two-stage refiner graphs, all covered in the ComfyUI 2026 complete guide
- Parameter optimization: fix the seed, then change one setting at a time to learn what each does
- Newer models: once SDXL feels comfortable, our Flux in ComfyUI guide walks through the fp8 and GGUF workflows
If the node graph feels like too much and you mainly want fast SDXL images, Fooocus runs the same checkpoints behind a simple form. Our Fooocus vs ComfyUI vs Automatic1111 comparison explains when each frontend is the right pick.
FAQ
What is the best SDXL model for ComfyUI in 2026?
Juggernaut XL is the best all-round realistic SDXL checkpoint for ComfyUI, with RealVisXL a close second for product and portrait work. For anime use AAM XL AnimeMix, and for one model that covers every style use DreamShaper XL.
Where do SDXL models go in ComfyUI?
Checkpoints go in ComfyUI/models/checkpoints, standalone VAE files in ComfyUI/models/vae, and LoRAs in ComfyUI/models/loras. Refresh the interface after copying a file so the dropdowns pick it up.
Do I need the SDXL refiner in ComfyUI?
No. Community fine-tunes such as Juggernaut XL and DreamShaper XL produce finished images without a refiner. The two-stage base plus refiner workflow is only worth building if you run the plain SDXL base model and want extra fine detail.
How much VRAM does SDXL need in ComfyUI?
SDXL at 1024x1024 runs on 8 GB of VRAM, and 12 GB or more gives you room for LoRAs, ControlNet and larger batches. On 6 GB cards, the --lowvram flag and fp8 loading make it possible but slow.
Why are my SDXL images black in ComfyUI?
Black images are almost always caused by the SDXL VAE overflowing in half precision. Load a fixed fp16 SDXL VAE through a Load VAE node, or launch ComfyUI with the --fp32-vae flag.
What sampler and CFG should I use for SDXL?
Start with DPM++ 2M Karras at 20 to 30 steps and a CFG between 5 and 8, then follow the recommendations on the specific model card. Turbo and Lightning checkpoints need far fewer steps and a much lower CFG.
Conclusion
SDXL in ComfyUI offers powerful image generation capabilities. Put a good fine-tune such as Juggernaut XL or DreamShaper XL in models/checkpoints, generate at a one-megapixel canvas with 20 to 30 steps, and add a fixed VAE only if your colors look off. With proper setup and understanding, you can create stunning AI-generated artwork efficiently.
Top comments (0)