Quick Answer: For AI image upscaling in 2026, Real-ESRGAN is the best value at $0.003 per image — good enough for the large majority of product photos. Topaz Image Upscale ($0.10) and Riverflow 2.0 RefSR ($0.20) are the quality leaders, GFPGAN Face Pro ($0.025) is best for faces, and Crystal Upscaler ($0.05) is the balanced mid-tier pick. The cost spread across upscalers is over 60x, so matching tool to job matters more than picking the "best" one.
The real cost of AI upscaling
Most upscaler comparisons rank tools by subjective quality and quote consumer subscription prices. This one uses the actual per-image inference costs we pay to run these models in production on PixelMotion.
| Model | Cost per image | Best for |
|---|
| Real-ESRGAN | $0.003 | General upscaling — the default | | Magic Image Refiner | $0.020 | Refinement + detail recovery | | GFPGAN Face Pro | $0.025 | Faces and portraits | | Recraft Crisp | $0.030 | Clean, sharp commercial output | | Crystal Upscaler | $0.050 | Balanced quality/cost mid-tier | | Topaz Image Upscale | $0.100 | Print and large-format quality | | Google Nano Banana Pro | $0.150 | High-end generative enhancement | | Riverflow 2.0 RefSR | $0.200 | Maximum-fidelity reference-based SR |
The spread is 66x — from $0.003 to $0.200 per image. On a 500-image catalog that is the difference between $1.50 and $100.
Which upscaler should you use?
Real-ESRGAN — $0.003 — the default answer
For the overwhelming majority of e-commerce and marketing work, Real-ESRGAN is the correct choice, and it is not close on price. At a third of a cent per image you can upscale an entire catalog for the cost of a coffee.
It is a well-established, general-purpose super-resolution model. It handles product photos, textures, and general imagery reliably. It is not the sharpest option on difficult material, but "not the sharpest" rarely matters at web display sizes.
Use it for: bulk catalog upscaling, web images, social content, anything at scale.
GFPGAN Face Pro — $0.025 — anything with people
General upscalers are trained on general imagery, and faces are where they fail most visibly — an upscaled face that's slightly wrong reads as uncanny immediately. GFPGAN is purpose-built for face restoration.
Use it for: portraits, team photos, lifestyle imagery with visible faces, UGC-style content.
Crystal Upscaler — $0.05 — the balanced middle
Roughly 16x the cost of Real-ESRGAN and half the cost of Topaz. The sensible pick when Real-ESRGAN isn't holding up on a specific image but you don't need print-grade output.
Topaz Image Upscale — $0.10 — print and large format
Topaz has the strongest reputation in the space for good reason: it holds detail under aggressive upscaling better than general models. That matters when the output will be printed, displayed large, or heavily cropped.
Use it for: print collateral, packaging, billboards, hero images, anything viewed at size. Don't use it for: 500 thumbnails. That is a $50 run that Real-ESRGAN does for $1.50.
Riverflow 2.0 RefSR — $0.20 — maximum fidelity
The most expensive option in the lineup, and reference-based rather than blind super-resolution. Reserve it for the small number of images that genuinely justify a 66x cost premium over the default.
The decision table
| Situation | Model | Cost |
|---|
| 500-image catalog | Real-ESRGAN | $1.50 total | | Product hero image | Topaz Image Upscale | $0.10 | | Photos containing faces | GFPGAN Face Pro | $0.025 | | Real-ESRGAN wasn't enough | Crystal Upscaler | $0.05 | | Print / packaging | Topaz or Riverflow RefSR | $0.10–$0.20 |
The mistake that wastes the most money
Running an entire catalog through a premium upscaler. A 500-image catalog costs $1.50 through Real-ESRGAN and $100 through Riverflow 2.0 RefSR — and for images displayed at 600px on a product grid, essentially nobody can tell.
The inverse mistake is cheaper but also real: upscaling your one print-bound hero image with the $0.003 model.
Match the tool to where the image will be seen. That single decision governs upscaling economics almost entirely.
Upscaling vs. enhancement — an important distinction
"Upscaling" (increasing resolution while preserving detail) and "enhancement" (changing or generating detail) are different operations, and several models in the table do the latter.
Google Nano Banana Pro ($0.15) and the Flux/Recraft/Ideogram families are generative — they can invent detail that wasn't in the source. That is powerful for creative work and actively dangerous for product photography, where invented detail means you are showing customers a product that doesn't exist.
For product images that a customer will buy from, prefer true super-resolution (Real-ESRGAN, Topaz, Crystal) over generative enhancement.
How we got these numbers (and what they do not include)
Methodology, stated plainly so you can judge the figures: every cost above is the per-second rate we are charged to run that model through our inference providers — Replicate and fal — read directly from our production model configuration. They are what a real platform pays per generation, not scraped vendor marketing pages.
Two caveats that matter when you compare this table to other sources:
- These are aggregator prices, not vendor-direct list prices. Running Runway through Replicate is not the same commercial arrangement as buying Runway credits from Runway. Direct-from-vendor pricing, enterprise agreements, and credit-bundle discounts can all differ — sometimes substantially. If you see a different number elsewhere, that is the most likely reason, and it does not make either figure wrong.
- This is generation cost only. It excludes storage, bandwidth, post-processing, and — the big one — failed or discarded generations. In practice your effective cost per usable clip is higher than the sticker rate, because you rarely keep the first output. If it takes three attempts to get a shot you ship, your real cost is roughly 3x the number in this table.
Provider pricing also changes. Treat this as a dated snapshot (July 2026) rather than a permanent quote, and verify before making a large commitment.
Frequently Asked Questions
What is the best AI image upscaler in 2026?
For most work, Real-ESRGAN at $0.003 per image offers by far the best value and is sufficient for web and e-commerce use. Topaz Image Upscale ($0.10) and Riverflow 2.0 RefSR ($0.20) lead on quality for print and large-format output, and GFPGAN Face Pro ($0.025) is the specialist choice for faces.
How much does AI image upscaling cost?
Between $0.003 and $0.20 per image depending on the model — a 66x spread. Real-ESRGAN is $0.003, Magic Image Refiner $0.02, GFPGAN Face Pro $0.025, Crystal Upscaler $0.05, Topaz $0.10, and Riverflow 2.0 RefSR $0.20. A 500-image catalog therefore costs anywhere from $1.50 to $100.
Is Topaz worth the price for upscaling?
For print, packaging, or any image displayed at large size, yes — it holds detail under aggressive upscaling better than general-purpose models. For a web catalog viewed at thumbnail or grid size, no: Real-ESRGAN produces comparable perceived quality at roughly one thirtieth the cost.
What is the best upscaler for faces?
GFPGAN Face Pro ($0.025 per image). General-purpose upscalers are trained on general imagery and fail most visibly on faces, where small errors read as uncanny immediately. A face-specific restoration model avoids that failure mode.
Does upscaling add detail that wasn't there?
It depends on the model, and the distinction matters. True super-resolution models (Real-ESRGAN, Topaz, Crystal) reconstruct resolution from what's present. Generative models (Nano Banana Pro, Flux, Recraft families) can invent detail. For product photography, prefer true super-resolution — invented detail means showing customers something the product doesn't actually have.
Upscale and enhance product photos with PixelMotion — 29 image models, free tier to start.
Try PixelMotion
Enhance your first product photo with AI