SUPIR Image Restoration and Upscaling: Complete ComfyUI Guide 2025

SUPIR represents a new generation of AI upscaling that goes beyond simple enlargement. It combines semantic understanding with image restoration, intelligently reconstructing missing details rather than just interpolating pixels.

Quick Answer: SUPIR (Scaling Up to Photo-Realistic Image Restoration) is an AI model that upscales and restores images by understanding their content. Unlike Real-ESRGAN which focuses on pattern matching, SUPIR uses semantic understanding to generate appropriate details for faces, textures, and objects.

What Makes SUPIR Different?

Traditional upscalers like Real-ESRGAN use pattern recognition to enhance images. They excel at clean inputs but struggle with degraded images. SUPIR takes a different approach:

Semantic Understanding: SUPIR understands what's in the image—faces, textures, objects—and generates details appropriate to that content.

Restoration + Upscaling: Instead of just enlarging, SUPIR first restores quality, then upscales. Compression artifacts, noise, and blur are addressed alongside resolution.

Generative Enhancement: Where details are missing, SUPIR generates plausible details rather than guessing from neighboring pixels.

SUPIR restores details that traditional upscalers cannot recover

When to Use SUPIR vs Real-ESRGAN

SUPIR is computationally heavier but produces superior results on degraded inputs.

Installing SUPIR in ComfyUI

Step 1: Install Custom Nodes

cd ComfyUI/custom_nodes
git clone https://github.com/kijai/ComfyUI-SUPIR

Step 2: Download Models SUPIR requires several model files:

• SUPIR-v0Q.ckpt (quality focused)
• SUPIR-v0F.ckpt (fidelity focused)
• Place in ComfyUI/models/checkpoints/SUPIR/

Step 3: Get Supporting Models SUPIR uses SDXL components:

• CLIP models for text encoding
• VAE for image encoding/decoding

Step 4: Check VRAM Requirements SUPIR needs 12GB+ VRAM for comfortable operation. 24GB recommended for high resolutions.

Understanding SUPIR Variants

SUPIR-v0Q (Quality): Prioritizes visual quality. May take creative liberties with details. Best for artistic enhancement where exact accuracy isn't critical.

SUPIR-v0F (Fidelity): Prioritizes accuracy to the original. Less aggressive enhancement but maintains original content better. Best for restoration where authenticity matters.

SUPIR Workflow Basics

A typical SUPIR workflow:

1. Load Image: Your degraded source image
2. SUPIR Encode: Prepares image for processing
3. SUPIR Condition: Sets parameters and optional text guidance
4. SUPIR Sample: The actual upscaling/restoration
5. SUPIR Decode: Converts back to image
6. Save: Output your enhanced image

Key parameters:

• Scale: Upscaling factor (2x, 4x, etc.)
• Steps: More steps = better quality, longer processing
• CFG Scale: How closely to follow guidance
• Positive Prompt: Optional text to guide restoration

Using Text Prompts for Guidance

SUPIR accepts text prompts that guide restoration:

Portrait Enhancement:

high quality photograph, detailed skin texture, sharp eyes,
professional photography, studio lighting

Landscape Restoration:

sharp landscape photograph, detailed foliage, clear sky,
professional nature photography, high resolution

General Quality:

high quality, sharp details, professional photography,
noise-free, artifact-free

Negative prompts help avoid artifacts:

blurry, noisy, jpeg artifacts, low quality, pixelated

SUPIR can produce 8K output from low-resolution inputs

Two-Stage Upscaling Workflow

For maximum quality, combine SUPIR with traditional upscalers:

Stage 1: SUPIR Restoration (2K)

1. Input degraded image
2. SUPIR restores and upscales to 2K
3. Output is clean but moderate resolution

Stage 2: ESRGAN Enhancement (8K)

1. Input SUPIR's 2K output
2. Apply 4x Foolhardy Remacri upscaler
3. Output is detailed 8K image

This methodical approach ensures each stage contributes to quality.

Optimizing SUPIR Performance

SUPIR is resource-intensive. Optimize with:

Memory Management:

• Enable attention slicing
• Use tiled processing for large images
• Close other GPU applications

Quality/Speed Balance:

• Fewer steps (20-25) for drafts
• More steps (40-50) for finals
• Lower CFG for speed, higher for quality

Tiling for Large Images:

• Process in tiles for images over 2048px
• Overlap tiles to avoid seams
• Blend edges in post-processing

Common Issues and Solutions

Issue: Out of VRAM Solution: Enable tiling, reduce resolution, or use --lowvram flag

Issue: Over-enhancement, uncanny look Solution: Use F variant instead of Q, reduce steps or CFG

Issue: Seams in tiled output Solution: Increase tile overlap, use blending in post-processing

Issue: Wrong details generated Solution: Add specific prompts guiding what details should look like

Issue: Processing too slow Solution: Reduce steps, lower resolution first pass, use FP16

SUPIR for Old Photo Restoration

SUPIR excels at restoring old photographs:

1. Scan Quality: Start with the best scan possible
2. Pre-Processing: Remove dust and major damage in Photoshop
3. SUPIR F Pass: Restore while maintaining fidelity
4. Color Correction: Fix fading and color shifts
5. SUPIR Q Pass: Optional enhancement for additional detail

The combination of F then Q variants often produces the best old photo restorations.

Batch Processing Considerations

For multiple images:

• Queue images in ComfyUI's batch system
• Use consistent settings for series
• Monitor VRAM usage across batch
• Consider overnight processing for large sets

Frequently Asked Questions

How does SUPIR compare to Topaz Gigapixel?

SUPIR offers similar quality with more control through prompts. Topaz is simpler but costs money. SUPIR runs locally for free but requires more VRAM.

Can SUPIR restore completely destroyed images?

SUPIR can generate plausible content for heavily degraded areas, but it's generating, not restoring. Results may not match the original.

What's the maximum upscale factor?

Practically, 4x is the limit for quality. Beyond that, run multiple passes (2x then 2x).

Does SUPIR work with video?

Frame-by-frame is possible but lacks temporal consistency. Dedicated video upscalers are better for motion content.

Why is SUPIR so slow?

It uses a full diffusion process, not just feed-forward networks. Quality requires iteration.

Conclusion

SUPIR represents the future of intelligent image upscaling—combining restoration and enhancement with semantic understanding. For degraded images where traditional upscalers fail, SUPIR can produce remarkable results.

The trade-off is computational cost. If you have the hardware (12GB+ VRAM) and patience for longer processing, SUPIR delivers quality that simpler tools cannot match.

For best results, understand when to use SUPIR versus faster alternatives, and consider two-stage workflows that leverage multiple tools' strengths.