How to Upscale Images with SUPIR in ComfyUI (Step-by-Step Tutorial)

In this tutorial, you’ll learn how to use the SUPIR (Super-resolution via Implicit Regularizers) node suite inside ComfyUI to upscale images with sharp detail and minimal artifacts. SUPIR leverages advanced latent-space encoding and diffusion techniques to transform low-resolution inputs into high-quality outputs, making it ideal for photography, digital art, and restoration workflows.

This walkthrough provides a downloadable workflow, detailed setup instructions, and recommended settings so you can get up and running in minutes.

SUPIR is a custom ComfyUI extension designed for first-stage diffusion super-resolution:

• High-fidelity upscaling: Retains fine textures and edges.

• Latent-space processing: Works on compressed representations for speed.

• Adjustable noise/fidelity: Control the balance between sharpness and smoothness.

Where standard upscalers (bilinear, lanczos) struggle with noisy or blurry inputs, SUPIR’s learned regularizers produce crisp, lifelike results without over-sharpening.

Before diving in, ensure you have the following:

1. ComfyUI v1.16.0+ installed locally or on a cloud server.

2. A modern GPU with at least 8 GB VRAM (12 GB+ recommended).

3. Download the SUPIR Checkpoint and SDXL Checkpoint

   Model Type	Checkpoint File Name	Download Link
   SUPIR Checkpoint	SUPIR-v0F_fp16.safetensors	Hugging Face
   SDXL Model Checkpoint	juggernautXL_v9Rdphoto2Lightning.safetensors	Civitai

To organize your files, place both the SUPIR and SDXL checkpoints under your ComfyUI models/checkpoints/ directory:

ts
1📂ComfyUI/
2└── 📂models/
3    └── 📂checkpoints/
4        ├── SUPIR-v0F_fp16.safetensors
5        └── juggernautXL_v9Rdphoto2Lightning.safetensors

To get started, you’ll first import our prebuilt workflow into ComfyUI—this gives you the complete node graph, including SUPIR encode/decode, resizing, and comparison nodes.

First, download the annotated workflow JSON from our CDN:

• SUPIR-upscale-next-diffusion.json

Once you have the file:

1. Drag-and-drop it onto the empty ComfyUI canvas. You’ll instantly see a cluster of nodes representing the end-to-end SUPIR pipeline. Some nodes will appear in red, indicating missing custom nodes. Let's fix that!

   

2. In the top right corner, click on 'Manager', then select 'Install Missing Custom Nodes'. From there, install the missing nodes by clicking 'Install'. The following nodes are missing:

   

   • ComfyUI-SUPIR

   • ComfyUI-KJNodes (for color matching)

   • rgthree-comfy (for side-by-side comparisons)

   • ComfyUI_essentials (for advanced resizing)

3. Restart or refresh ComfyUI. Once reconnected, the red outlines will be gone, confirming they’re ready to run.

With the workflow loaded and dependencies handled, you’re now set to dive into fine-tuning SUPIR settings for your own images.

Tuning SUPIR effectively means understanding how each parameter shapes the upscaling process. Below, we break down the most important controls—and explain why you might tweak them:

By default, SUPIR takes your latent encoding through a diffusion pass that balances guidance strength, noise injection, and decoding granularity. These three axes correspond to:

SUPIR Sampler Node Settings

• Steps (10–30):
  SUPIR’s “steps” determine how many diffusion iterations it performs.

  • Fewer than 15 speeds up the upscale but can leave fine details under-refined.

  • More than 20 adds extra polish at the cost of longer runtimes.

• CFG Scale Start / End (e.g. 2.5 → 1.5):
  This ramp controls how forcefully SUPIR follows your prompts over the diffusion timeline.

  • A higher start value locks in strong guidance early, preserving macro structure.
  • A lower end value allows for natural smoothing as it finishes.

• EDM_s_churn & s_noise (e.g. 5 & 1.002):
  These tweak how much “shake” (noise) is injected at each step.

  • Higher churn creates more texture and grit.
  • Slight noise boosts sharpen edges but can introduce grain if overdone.

SUPIR Decode Node Settings

Decoder tile size settings range from 512–768.** Choosing your tile size directly affects VRAM usage and output uniformity.

• 768 is ideal on 12 GB+ cards for artifact-free stitching.

• 512 on an 8 GB GPU avoids out-of-memory errors with only a minor stitch-line trade-off.

Pro Tip: After finding a configuration that hits your sweet spot, save it as a ComfyUI preset. That way, every new project starts with your optimized baseline.

Now that you’ve got the workflow and know which knobs to twist, let’s execute an upscale from start to finish. The process is linear—load, encode, sample, decode, compare—but each stage informs the next.

Begin by wiring up your nodes:

1. Load Image
   Simply drag or select your low resolution image. ComfyUI handles any format (PNG, JPEG, TIFF).

2. Image Scaling (optional)
   If your input is very small (<512 px), upsample to a modest size (e.g., by 3) before entering latent space. This often gives SUPIR more room to work.

3. Load Checkpoint & SUPIR model loader v2
   Select both your SUPIR checkpoint and the SDXL model. This node outputs all necessary model and VAE handles.

4. Run
   Hit the Run button and watch as your GPU churns through each step. Monitor the ComfyUI console for VRAM usage and any warnings.

Once complete, your output node holds the final SUPIR-upscaled image.

Low Quality Input Image

SUPIR Upscaled Output

Notice how the orange sports car’s paintwork regains sharp definition, the outdoor background details pop with clarity, and the low-quality JPEG artifacts are effectively minimized without introducing new noise.

Low Quality Input Image

SUPIR Upscaled Output

In this example, notice how the woman’s facial features are gently restored—the eyes gain definition, skin texture becomes more natural, and hair strands are clearly separated without harsh over-sharpening. Despite starting from a heavily compressed image, SUPIR recovers clarity and realism while preserving softness where appropriate, making it ideal for restoring old portraits or social media-quality selfies.

By merging SUPIR’s advanced regularizers with SDXL’s expressive capacity, ComfyUI becomes a powerhouse for super-resolution upscaling. With just a few clicks—loading the workflow, installing nodes via the Manager, and tweaking a handful of sliders—you can elevate any image from blurry to breathtaking. Download the workflow, fine-tune your presets, and let SUPIR redefine what’s possible in your creative pipeline.