Skip to content

Image Upscaling & Print Production Guide

Overview

This guide covers the complete workflow of upscaling images to print-ready resolution. It addresses three primary source types — AI-generated art, scanned physical artwork, and digital photography — with specific recommendations for each.

Goal: Produce the highest quality, most natural-looking enlarged image for fine art printing and Print on Demand (POD) production.

Philosophy: Upscaling is not magic. It works best when the source image is already high quality. The upscaler's job is to enlarge without introducing visible artifacts — not to invent detail that was never there.


Resolution Fundamentals

DPI, Pixels, and Print Size

DPI (Dots Per Inch) describes how many pixels fit into one inch of printed output. The same pixel count prints larger at lower DPI and smaller at higher DPI.

Formula: Required Pixels = Print Size (inches) x DPI

Minimum Resolution by Print Type

Print Type Minimum DPI Notes
Fine art paper print 300 Viewed up close — maximum detail needed
Canvas gallery wrap 150–200 Canvas texture hides minor imperfections
Large format poster (> A2) 150–225 Viewed from distance
Metal / sublimation print 300 High-contrast medium, detail matters
DTG garment (t-shirt) 150–300 Depends on platform and design complexity
Billboard / banner 20–72 Viewed from several meters

Common Print Size Pixel Requirements (300 DPI)

Print Size (inches) Print Size (cm) Pixels Required
8 x 10 20 x 25 2 400 x 3 000
11 x 14 28 x 36 3 300 x 4 200
16 x 20 41 x 51 4 800 x 6 000
18 x 24 46 x 61 5 400 x 7 200
24 x 36 61 x 91 7 200 x 10 800
30 x 40 76 x 102 9 000 x 12 000

POD Platform Minimum Requirements

Platform Min. Resolution Aspect Ratio Format
Fine Art America 6 000 px (long edge recommended) Flexible JPEG, PNG
Displate 2 900 x 4 060 px 1:1.4 JPG
Redbubble 4 500 x 5 400 px (poster) Varies by product PNG recommended
INPRNT 4 000 px (short edge recommended) Flexible JPEG, PNG, TIFF
Printful (poster) 300 DPI at print size Varies PNG recommended

Note: Platform requirements change. Always verify current specs on the platform's upload page before preparing files.


Source Types and Their Characteristics

AI-Generated Images

Typical native resolution:

Tool Native Output
DALL-E 3 (ChatGPT) 1 024 x 1 024 or 1 024 x 1 792
Stable Diffusion (SD 1.5) 512 x 512 (optimal)
Stable Diffusion XL (SDXL) 1 024 x 1 024 (optimal)
Midjourney v6+ Up to 2 048 x 2 048

Upscaling need: Almost always required for print. A 1 024 x 1 024 image at 300 DPI prints only 3.4 x 3.4 inches (8.7 x 8.7 cm) — too small for most products.

Advantages for upscaling:

  • Clean digital source — no film grain, sensor noise, or scanning artifacts
  • Colors stay within sRGB gamut — no out-of-gamut surprises
  • No specular reflection or physical texture issues
  • Consistent quality across the image

Watch for:

  • AI artifacts in hands, text, fine geometric patterns
  • Inconsistent fine detail that upscaling may amplify
  • Some AI generators apply JPEG compression — save/download in highest quality

Scanned Physical Artwork

Typical resolution: 600–1 200 DPI from a flatbed scanner (e.g., Epson V600)

Upscaling need: Usually minimal if scanned at 600+ DPI. A 600 DPI scan of an A4 original (8.3 x 11.7 inches) produces a 4 980 x 7 020 pixel image — sufficient for fine art prints up to approximately 16 x 23 inches at 300 DPI without upscaling.

Challenges:

  • Scanner light direction creates directional artifacts on textured surfaces
  • Specular reflection from glossy/metallic media (gel pens, metallic inks)
  • Dust, hair, and surface debris appear as bright spots on dark media
  • Auto-exposure may struggle with predominantly dark artwork

See the dedicated Scanning Artwork on Dark Surfaces section below.

Digital Photographs

Typical resolution: 12–60+ megapixels depending on camera

Upscaling need: Rarely needed for the full image — more commonly needed when cropping significantly reduces usable resolution.

Watch for:

  • Sensor noise (especially high ISO) — upscaling amplifies noise
  • Lens softness at edges — upscaling cannot recover true detail
  • JPEG compression artifacts from in-camera processing
  • Motion blur — upscaling sharpens edges but cannot undo motion blur

Phone Photos of Physical Artwork

Typical resolution: 12–200 MP (modern smartphones)

This is the worst source type for reproduction. Problems include:

  • Uneven lighting and shadows
  • Perspective distortion (keystoning)
  • Color temperature inconsistency
  • Lens distortion at edges
  • Reflection from glossy surfaces
  • Compression artifacts

Recommendation: Always use a flatbed scanner for flat artwork. Reserve phone photography only for initial concept documentation or when the artwork exceeds the scanner bed.


Upscaling Tools

What: Free, open-source, GPU-accelerated image upscaler using Real-ESRGAN and related models.

Why recommended:

  • Completely free and open-source
  • Runs locally — nothing leaves the machine
  • GPU-accelerated via Vulkan (works with NVIDIA, AMD, Intel GPUs)
  • Multiple AI models included
  • Simple interface for quick processing
  • PNG output (lossless) by default

System Requirements:

Component Minimum Recommended
GPU Vulkan-compatible (GTX 1060 / RX 580) GTX 1070+ / RTX series
VRAM 4 GB 6–8 GB
RAM 8 GB 16 GB
Storage ~500 MB for app + models SSD recommended

Key Settings:

Setting Recommendation Notes
Output Format PNG Lossless — always use for intermediate files
Compression 0 Default; lossless PNG compression
Scale Factor 4x Maximum useful single-pass factor
Tile Size 512 (default) Reduce to 256 or 128 if VRAM errors occur
Double Upscayl Use cautiously Chains two 4x passes (= 16x total); inspect at 100% zoom

Practical Limits:

  • 4x is the sweet spot for most upscaling tasks
  • Beyond 4x (using Double Upscayl), the AI begins inventing detail — results become speculative
  • Always inspect the output at 100% zoom before committing to production
  • Large source images (8 000+ px) may cause VRAM overflow — reduce tile size or process in sections

ComfyUI Integrated Upscaling

For users already working in ComfyUI for image generation, upscaling can be integrated directly into the generation workflow.

Core Nodes:

Node Purpose
UpscaleModelLoader Loads an ESRGAN model from models/upscale_models/
ImageUpscaleWithModel Applies the loaded model to upscale an image
Ultimate SD Upscale (custom node) Tile-based upscaling with optional img2img refinement

Model Installation: Place .pth model files in ComfyUI/models/upscale_models/. Models are available from OpenModelDB and Civitai.

Two Approaches:

1. Simple Model Upscale (non-generative):

  • Load model → feed image → get upscaled result
  • Fast, deterministic, same result every time
  • No KSampler needed
  • Best for: final production upscaling of a finished image

2. Generative Tiled Upscale (img2img refinement):

  • Upscales with ESRGAN, then runs a tiled img2img pass to add AI-generated detail
  • Uses KSampler + ControlNet Tile for fidelity
  • Slower, non-deterministic (varies between runs)
  • Denoise: 0.3–0.5 (lower = more faithful to original; higher = more creative freedom)
  • Best for: adding fine detail to images that lack texture at higher resolution

Caution: Generative upscaling changes the image. New detail is invented by the model, not recovered from the source. Always compare with a non-generative upscale to verify the result is acceptable.

Other Notable Tools

Tool Type Cost Notes
Topaz Gigapixel AI Commercial, local ~$199 (one-time) Widely regarded as high quality; proprietary models
Let's Enhance Cloud-based From $9/month Convenient; images leave your machine
GIMP + plugins Open-source Free Limited AI upscaling capabilities
ImageMagick CLI Free Non-AI; bicubic/Lanczos only — useful for precise resizing, not AI upscaling
Waifu2x Open-source Free Anime/illustration focused; predecessor to modern ESRGAN models

Note on cloud tools: Any cloud-based upscaling service processes the image on their servers. For unpublished or sensitive artwork, local processing (Upscayl, ComfyUI) is preferable.


Model Selection

Upscayl Models — Detailed Comparison

Model Best For Characteristics Watch For
Real-ESRGAN (General Photo) Photographs, mixed real-world images Balanced; trained on realistic degradation patterns May smooth fine artistic textures
Real-ESRGAN (Fast) Quick previews, batch processing Faster but lower quality than standard Not for final production
4x-UltraSharp AI-generated art, illustrations, text Aggressive detail reconstruction; sharpest edges Can over-sharpen; may create artifacts on organic textures (grass, skin)
Remacri Best all-rounder; mixed content Balanced detail vs. artifact avoidance; good text/screenshot handling Slightly less sharp than UltraSharp on clean sources
Ultramix Balanced Urban scenes, architecture Good structural detail preservation Significantly slower (5–6x) than other models
x4plus-anime Anime, manga, cel-shaded illustration Clean line handling, flat color region preservation Will smear photographic skin texture

Model Selection Decision Tree

Start here → What is the source image?

  1. AI-generated illustration / digital art → Try 4x-UltraSharp first. If over-sharpened → Remacri
  2. Photograph (digital camera)Real-ESRGAN or Remacri
  3. Scanned line art / ink drawing4x-UltraSharp
  4. Scanned painting / textured artworkRemacri (preserves texture better)
  5. Anime / manga / flat-color illustrationx4plus-anime
  6. Urban / architectural photoUltramix Balanced (if time permits)
  7. Mixed content / unsureRemacri (safest default)

Advanced: Model Chaining

Some workflows chain two different models for potentially better results:

  • Real-ESRGAN 2x → then UltraSharp 2x = combines ESRGAN's natural texture with UltraSharp's edge definition
  • Any model 4x → then Lanczos downscale to target size = upscale higher than needed, then downsample for extra sharpness

Uncertain: Model chaining results are subjective and content-dependent. Always A/B test against a single-pass 4x upscale before adopting a chained workflow. The quality improvement is not guaranteed and adds processing time.


Scanning Artwork on Dark Surfaces

This section addresses the specific challenge of scanning artwork created on black cardboard or paper using gel pens, metallic inks, colored pencils, or markers — common in certain illustration styles.

Why Dark-Surface Artwork Is Difficult to Scan

1. Specular Reflection

Gel pens and metallic inks have glossy, reflective surfaces. A flatbed scanner's light source hits the artwork at a fixed angle. Where the ink's surface angle aligns with the light-sensor path, the scanner captures a bright specular highlight instead of the actual color. This appears as white or washed-out spots in the scan.

2. Auto-Exposure Confusion

Most scanner software auto-exposure algorithms assume a predominantly light image. A mostly-black artwork causes the software to dramatically increase exposure, blowing out the highlights (the actual artwork) and lifting the black background to noisy grey.

3. Scanner Lid Reflection

The standard white scanner lid reflects light back through the artwork. On dark paper, this is usually not a problem (unlike with film scanning), but on semi-transparent or thin dark paper, it can contaminate the black with a milky appearance.

Scanner Settings for Dark Artwork

Use Professional / Manual Mode — never Auto mode.

Setting Value Reason
Document Type Reflective Standard for opaque artwork
Image Type 48-bit Color Maximum color depth for post-processing latitude
Resolution 600 DPI Sufficient for most reproduction; 1 200 DPI only if final print will be significantly larger than the original
Color Restoration OFF Designed for faded photos — will distort intentional color choices
Dust Removal (ICE) OFF May remove intentional fine marks; clean manually in post
Unsharp Mask OFF Apply sharpening in post-processing where you have more control
Auto Exposure OFF Set manually to prevent black level corruption
Backlight Correction OFF Not applicable to opaque artwork

Manual Exposure Adjustment:

  • Set the black point so that the cardboard background is deep black but not clipped to pure #000000 — you want to preserve the paper's subtle texture
  • Set the white point / highlight so that the brightest gel pen strokes are captured without clipping

Techniques for Reducing Specular Reflection

1. Cork Buffer Method (Most Effective for Flatbed)

Place thin cork strips (1–2 mm) at the corners between the artwork and the scanner glass. This creates a small air gap that changes the angle of light interaction, reducing direct specular reflection from glossy media.

Source: Muddy Colors — A Quick Trick for Scanning Textured Surfaces

2. Black Backing Instead of White Lid

Replace or cover the scanner's white lid with matte black paper or card. This eliminates reflected light bouncing back through the glass from the lid, reducing overall light contamination on dark artwork.

3. Multi-Pass Scanning (180° Rotation)

  1. Scan the artwork in its normal orientation
  2. Rotate the artwork 180° on the scanner bed
  3. Scan again with identical settings
  4. In Photoshop: layer both scans, rotate one 180° to align, then blend (use lighter/darker blending or manual masking) to average out directional lighting artifacts

Source: Muddy Colors — How I Scan and Assemble My Work

4. Multi-Pass for Metallic Media

For artwork with both matte and metallic/glossy elements:

  1. Scan once with standard exposure (optimized for matte areas)
  2. Scan again with reduced exposure (optimized for metallic highlights)
  3. Combine in Photoshop using layer masks — matte areas from scan 1, metallic areas from scan 2

5. Photography Alternative (For Highly Reflective Media)

When scanning cannot capture metallic effects:

  • Set up two lights at 45° angles to the artwork
  • Use polarizing filters on both lights AND the camera lens (cross-polarization)
  • This eliminates specular highlights while preserving color
  • Requires careful color calibration against a reference target

Note: Cross-polarization photography requires specialized equipment (polarizing gels for lights, CPL filter for camera) and is more complex than scanning. Only necessary when flatbed scanning consistently fails to capture metallic effects.

Physical Preparation Before Scanning

  1. Allow ink/paint to fully cure (gel pens: 24+ hours for metallic varieties)
  2. Clean scanner glass with lint-free microfiber cloth and electronics-safe glass cleaner
  3. Wipe artwork gently with anti-static cloth to remove dust and fibers
  4. On black paper/card, every speck of dust appears as a bright white dot — thorough cleaning is critical
  5. Handle artwork by edges only — fingerprints show on both the glass and glossy ink surfaces

Post-Scan Processing

Step Tool Notes
Dust/debris removal Photoshop — Healing Brush, Clone Stamp Manual removal; auto dust removal may damage fine lines
Black level adjustment Photoshop — Levels / Curves Bring background to deep black while preserving paper texture
Color correction Photoshop — Curves per channel Correct any color cast from scanner light source
Sharpening Photoshop — Unsharp Mask or Smart Sharpen Apply at final output resolution, not before upscaling
Save TIFF 16-bit or PNG Lossless format for archival master file

Color Space Considerations

Workflow Stage Color Space Notes
AI image generation sRGB All major AI generators output sRGB
Scanning (archival) Adobe RGB (if scanner supports) Wider gamut preserves more color data
Editing / post-processing Adobe RGB or ProPhoto RGB Edit in the widest space practical
Upload to POD platform sRGB Most POD platforms expect sRGB; they handle CMYK conversion internally
Upload to professional print lab Check lab requirements Some accept Adobe RGB or ICC-profiled TIFF

Important: Do NOT manually convert to CMYK for POD platforms. Their RIP (Raster Image Processor) software uses custom ICC profiles calibrated to their specific printers and papers. Your home CMYK conversion will be less accurate than theirs.

Gamut Limitations — What Cannot Be Printed

Certain colors visible on screen or in physical media cannot be reproduced in print:

Medium Gamut Issue
Neon / fluorescent gel pens Colors exceed sRGB and CMYK gamut — will appear duller in print
Metallic inks (gold, silver) Reflective quality cannot be reproduced with standard CMYK ink on paper
Holographic / iridescent media Angle-dependent color impossible in static print
Extreme saturated colors (electric blue, hot pink) sRGB can display these; CMYK often cannot match

Compensation strategies:

  • Increase Vibrance +5–10% in post-processing to compensate for print dullness
  • Consider metallic paper or metal substrate prints for artwork originally created with metallic media
  • Accept that some physical media effects are inherently non-reproducible — communicate this clearly in product descriptions if necessary

Optimization by Print Medium

Paper Prints (Giclée, Fine Art)

  • Paper choice matters: Matte / rag paper absorbs light → deep, bársonyos fekete effect
  • Avoid glossy paper for predominantly dark images: High ink coverage on glossy paper causes visible print head banding, fingerprint collection, and distracting reflections
  • Heavy paper (250+ g/m²): Required for high ink coverage — prevents warping/cockling
  • Recommended paper types for dark artwork: Somerset Velvet, Hahnemühle Photo Rag, Canson Infinity Rag Photographique

Note: Paper choice is only relevant when printing yourself or using a custom print lab. POD platforms select their own paper stock; your control is limited to choosing between "matte" and "glossy" options (always choose matte for dark backgrounds).

Metal Prints (Sublimation)

Metal prints are ideal for dark-background artwork because:

  • Sublimation dye is infused into the metal coating under heat — no ink sitting on the surface
  • The metal substrate itself provides depth and luminosity
  • No ink saturation / warping problems
  • The reflective metal surface naturally reproduces the "glow" effect of bright colors on dark backgrounds
  • Durable, scratch-resistant, waterproof

Technical requirements:

  • sRGB color space
  • 300 DPI at target print size
  • JPEG or PNG (check platform requirements)

Canvas Prints

  • Canvas texture naturally hides minor upscaling artifacts and printer imperfections
  • 150–200 DPI is generally acceptable for canvas (lower than paper)
  • Gallery wrap requires extra image area for the edges (typically 1.5–2 inches per side)
  • Dark backgrounds on canvas can appear slightly less deep than on paper or metal

Garments (DTG — Direct to Garment)

For dark-background artwork on garments:

  • Use transparent PNG — remove the black background entirely
  • The garment fabric provides the "black background"
  • This avoids the printer applying a thick ink layer for the background → no rubbery/stiff feel
  • If the background contains intentional texture/pattern (not solid black), leave it — but the garment should be black

Caution: If a solid digital black (#000000) background is left in the file and printed on a black garment, the printer will still lay down black ink on black fabric. The result is a visible, slightly different-shade rectangle with a stiffer texture that cracks after washing.


Quality Verification Checklist

After upscaling and before uploading to any platform:

Check How Pass Criteria
100% zoom inspection View at actual pixels in Photoshop No visible artifacts, halos, or "plastic" smoothing
Edge quality Zoom into sharp edges, text, fine lines Edges are clean without ringing or over-sharpening
Texture preservation Compare with source at similar zoom Paper grain, brush strokes, pencil texture maintained (not smoothed away)
Background uniformity Inspect dark/solid areas at 100% No introduced noise, pattern artifacts, or color banding
Color accuracy Compare with source side by side No color shift introduced by upscaling model
Resolution Check pixel dimensions and DPI Meets target platform's minimum requirements
File format Verify output format PNG for intermediate; platform-specific format for upload
File size Check file size is reasonable Unusually small may indicate compression; unusually large may cause upload issues

Test Print Protocol

Always order a proof before listing a product for sale.

  1. Order the smallest/cheapest size available on the platform
  2. Choose the same paper/material type as the production listing
  3. Evaluate in person: check blacks, color accuracy, sharpness, texture
  4. View at intended hanging/viewing distance
  5. If unsatisfactory, adjust and re-proof before listing

Common Pitfalls

Mistake Why It Happens Solution
Upscaling a JPEG multiple times Each save adds compression artifacts that compound Always work from the original source; upscale once from the best version
Using the wrong model Applying a photo model to illustrations (or vice versa) Match model to content type — see Model Selection section
Upscaling too aggressively 16x upscale on a low-res source Maximum 4x in a single pass; beyond that, quality degrades rapidly
Ignoring 100% zoom check Thumbnail looks fine but actual pixels show artifacts Always inspect at 100% before committing
Not ordering a test print Screen and print differ — always Budget for test prints as part of the production cost
Manual CMYK conversion for POD Trying to control what the printer does Upload sRGB; let the platform handle conversion
Leaving solid black BG on garments Assumes the printer "knows" the garment is black Export transparent PNG for dark garments
Over-sharpening after upscaling Stacking upscaler sharpening with post-process sharpening Apply sharpening only once, at the final step, at output resolution
Scanning at maximum DPI "just in case" 2 400+ DPI scans are enormous with no benefit for reflective media 600 DPI is sufficient for most artwork; 1 200 DPI maximum for paper originals

Quick Reference: Complete Workflow

AI-Generated Art → Print

Step Tool Details
1. Generate image Dall-E / ComfyUI + SDXL Request highest available resolution; specify desired textures in prompt
2. Quality check Photoshop Fix AI artifacts (hands, text, geometric errors)
3. Upscale Upscayl — 4x-UltraSharp or Remacri 4x factor; inspect at 100%
4. Color adjust Photoshop — Vibrance +5-10% Compensate for print dullness (optional, image-dependent)
5. Prepare variants Photoshop Full background PNG (wall art) + transparent PNG (garments)
6. Export PNG, sRGB Verify pixel dimensions meet platform requirements
7. Upload + test print Platform Always proof before listing

Scanned Artwork → Print

Step Tool Details
1. Prepare & scan Flatbed scanner, Professional Mode 600 DPI, 48-bit color, all auto corrections OFF
2. Clean Photoshop — Healing Brush Remove dust, debris; preserve intentional texture
3. Levels/Curves Photoshop Set black and white points; correct color cast
4. Upscale (if needed) Upscayl — Remacri or UltraSharp Only if print size significantly exceeds scan resolution
5. Sharpen Photoshop — Unsharp Mask Apply once at final resolution
6. Export TIFF 16-bit (archival) + PNG (upload) sRGB for POD; Adobe RGB for custom lab
7. Upload + test print Platform Always proof before listing

Digital Photo → Print

Step Tool Details
1. RAW develop RawTherapee / Lightroom Export at maximum resolution, 16-bit TIFF
2. Retouch Photoshop Standard post-processing
3. Upscale (if needed) Upscayl — Real-ESRGAN or Remacri Only if heavy crop reduced resolution below target
4. Sharpen Photoshop Output sharpening for print
5. Export TIFF (archival) + JPEG/PNG (upload) sRGB for POD
6. Upload + test print Platform Always proof before listing

Appendix: Pixel Dimension Calculator

To calculate required pixel dimensions for any print size:

Pixels = inches x DPI

Target At 150 DPI At 300 DPI
5 x 7" 750 x 1 050 1 500 x 2 100
8 x 10" 1 200 x 1 500 2 400 x 3 000
11 x 14" 1 650 x 2 100 3 300 x 4 200
16 x 20" 2 400 x 3 000 4 800 x 6 000
18 x 24" 2 700 x 3 600 5 400 x 7 200
20 x 30" 3 000 x 4 500 6 000 x 9 000
24 x 36" 3 600 x 5 400 7 200 x 10 800
30 x 40" 4 500 x 6 000 9 000 x 12 000
40 x 60" 6 000 x 9 000 12 000 x 18 000

For metric paper sizes at 300 DPI:

Paper Size Dimensions (mm) Pixels at 300 DPI
A5 148 x 210 1 748 x 2 480
A4 210 x 297 2 480 x 3 508
A3 297 x 420 3 508 x 4 961
A2 420 x 594 4 961 x 7 016
A1 594 x 841 7 016 x 9 933
A0 841 x 1 189 9 933 x 14 043

Sources and Further Reading


Last updated: 2026-08-11