New Supir Workflow ComfyUI

One node used to handle the whole job. Now it's been cracked open into a full suite... and the results are worth every extra connection you wire up.

SUPIR has been the favorite upscaler in the ComfyUI world for a while now. The original version? One big node. It worked. But the developers knew they could do better.

So they split it apart.

The new SUPIR suite breaks that single node into modular pieces... Model Loader, First Stage Denoiser, Encode, Conditioner, Sampler, Decode, and Color Match. Each one gives you finer control. Each one earns its spot on the canvas. And when you wire them together? The results get noticeably closer to what the original SUPIR demos promised.

Abigail from AI Fuzz walked through building the entire workflow from a blank ComfyUI project, and the process is more approachable than the node count might suggest.

Getting Started

First things first... installation. You grab the new SUPIR suite from GitHub and git clone it into your custom nodes folder. Or use the ComfyUI Manager to install it directly. Either way works.

Once installed, right-click anywhere on the canvas, go to Add Node, and the full SUPIR suite appears in the list. That's your toolbox.

Building the Graph

The workflow starts with the SUPIR Model Loader. This needs two things... an SDXL base checkpoint (like New Dawn XL or Juggernaut XL) and a dedicated SUPIR checkpoint. Pro tip: the V0F variant handles face detail significantly better. Drop both into your checkpoints folder.

From there, the build follows a logical chain:

1. Load Image → feeds into a Resize node (Abigail uses 1536×1536) 2. First Stage Denoiser → takes the resized image and the SUPIR VAE from the Model Loader 3. SUPIR Encode → receives the denoised latents and denoised image 4. SUPIR Conditioner → this is where it gets interesting

The Conditioner includes a text prompt. You're not just upscaling blindly... you're guiding the process. A prompt like "high quality, detailed photograph of female" tells the model what it's working with. This is text-guided upscaling, and it makes a real difference for specific subjects like faces and textures.

5. SUPIR Sampler → connects to the Conditioner's positive output and the Model Loader's SUPIR model output. Abigail sets steps to 100 and leaves the rest at defaults. 6. SUPIR Decode → keeps tile size at 512 7. Color Match → the final polish

That Color Match node matters more than you might think. After all that processing, colors can drift. The Color Match node (set to HM, though Reinhard and MBGD are solid options too) pulls the output back to the original image's color palette. Subtle. Essential.

The Results

Abigail ran two tests. The first... a somewhat blurry photo of a girl on a couch. The RGThree Image Comparer node tells the whole story with its interactive slider. Before: soft, fuzzy, lacking definition. After: sharp eyes, clean lip outlines, smoothed-out blur zones. Real improvement from a rough starting point.

Second test... a different image, same workflow. The results held consistent. Sharper details, reduced anti-aliasing artifacts, cleaner edges throughout.

Both outputs landed at 1536×1536. And if you need to push further? You can chain a second SUPIR pass or hand the output to SD Ultimate Upscale for another resolution bump.

Settings Worth Exploring

The beauty of the modular approach is the surface area for experimentation. You can adjust:

- Encoder tile size and encode D type - Sampler settings (steps, CFG, sampler type) - Color Match method (HM, Reinhard, MBGD) - Resize dimensions for the initial upscale target - Text prompts in the Conditioner for subject-specific guidance

Every one of those dials gives you a slightly different result. The old single node locked most of that away.

Why This Matters

The shift from one node to a full suite isn't just a technical upgrade. It's a philosophy shift. Instead of trusting a black box, you see every stage. You understand what each piece does. You can troubleshoot, tweak, and learn.

That's the thing about tools that grow with you... they don't just make better images. They make better practitioners.

The workflow stays compact enough to bolt onto existing ComfyUI pipelines. It's still the team favorite upscaler. It just has more muscle under the hood now.

If you've been running the old single-node SUPIR setup, this is your upgrade path. If you've never touched SUPIR before, this is a clean place to start. Either way... wire it up, run a blurry photo through it, and drag that comparison slider back and forth. The results speak louder than any tutorial. 💪

Go build something. Go make something sharper than it was before. Quietly working... one node at a time.

--- Source: https://www.youtube.com/watch?v=y2MzwsJUmb4