How Photographers Should Pick an AI Upscaler for Client Files
Guest post
Come on, darling, it's just a rock...
Every AI upscaler promises to "enhance" a file. That word hides two opposite operations. One tool recovers detail the sensor actually recorded. Another invents detail that looks sharp on a thumbnail but was never in the frame. When the file belongs to a client, that difference decides whether you can deliver it.
Three blind evaluations run by Everypixel's production team between June and August 2026 make the difference measurable. Each test used the same structure: 264 image pairs, 13 evaluators, three independent judgments per pair. In every test the open-weights restoration model SeedVR2 was compared against a different rival.
Restores vs rewrites: why it matters for a client file
A restoring upscaler tries to resolve structure that is present in the source but blurred at its resolution. A rewriting upscaler generates plausible new content. On a wedding guest's face, a product label or a shop sign, it changes the record of what you photographed.
The vendors say this openly. Magnific's API documentation describes its Creative path as "prompt-guided enhancement that can introduce or infer new detail", and it offers a separate Precision endpoint documented as "faithful upscaling". The blind panel in the Magnific test named the same mechanism as its main complaint: the model was "adding details and highlights that were not in the source", in the words of one art director.
What a blind panel is
In a blind paired test, each source image is upscaled by both tools. The two results are shown side by side with the model names hidden and the left/right position randomized. Evaluators pick A, B, or "equal" on each scale: detail recovery, edge sharpness and naturalness; the June test also scored color accuracy. Half of the 264 sources were real photographs and half were AI-generated, spread across six categories, from faces and crowds to text and signage.
How to read the win rates
A win rate is the share of pairs in which the majority of the three evaluators chose that model. Ties are counted separately, so the two win rates seldom add up to 100%. Two other numbers tell you how much to trust a split:
P(better) is a Bradley-Terry estimate: the probability that one model is preferred in a single comparison. A value of 0.50 means a coin toss.
Agreement (Gwet's AC1) shows whether evaluators agreed with each other. Near zero means the panel could not tell the outputs apart reliably, and the result is a direction, not a finding.
The Topaz row shows what a close result looks like. Agreement fell to AC1 0.06–0.12, against 0.42–0.71 for the same panel in June, so that report is labelled directional. The Magnific row is the opposite: on detail and naturalness agreement held near 0.5. The only weak scale there was edge sharpness (AC1 0.08, half the pairs tied). The authors read its 67% figure "as an orientation, not as a finding". The full write-up of a 264-pair blind test of two upscalers breaks every scale down by image category.
The face-detail caveat
Faces are where the simple story breaks. Against Magnific, SeedVR2 won 93% of face pairs on detail recovery, and Magnific won none. Against Topaz Wonder 3, the ranking reversed: Topaz was preferred on facial naturalness in 52% of pairs against 14%, and also led on face detail, 25% to 18%.
The reason is a trade-off between naturalness and fidelity. Evaluators said Topaz Wonder 3 renders skin as a continuous, even surface, but alters facial features more visibly. SeedVR2 preserves proportions and likeness more closely. It can over-render skin; one art director called it "lizard skin" on tight face crops. ByteDance's own model card warns that the models "tend to overly generate details on inputs with very light degradations."
When a paid tool still wins
The Topaz report recommends routing portrait work to Topaz Wonder 3 "where skin must read as unprocessed". Where likeness matters more than skin appearance, the preference goes back to SeedVR2. It adds that a 0.55 probability "is not grounds for migrating an existing workflow."
Magnific's case is narrower: graphics, illustration, abstract scenes and landscapes without people or fine text. At the time of that test (August 2026), Magnific listed no free tier, with plans from EUR 12 per month billed annually. SeedVR2 weights are released under Apache 2.0.
What these tests do not tell you
The Magnific data package does not record which endpoint was used. The behaviour matches the Creative path; Precision mode has not been benchmarked.
Topaz Wonder 3 was run in one configuration. Its authors note that parameter tuning could shift the result on any scale.
Only subjective scales were used. No reference metrics such as PSNR, SSIM or LPIPS were computed.
Sources came from production stock workflows. The authors warn against extrapolating to medical, scientific, forensic or archival restoration.
How to test on your own files
Pick 20–30 typical client files: portraits, groups, products, anything with text.
Run each through both tools at the same output size, and record the settings you used.
Have someone else randomise left/right and hide the names. Score detail, sharpness and naturalness as A, B or equal.
Count ties honestly. If most pairs tie, decide on cost and workflow, not quality.
Check faces for changed likeness and skin texture at 100%, not only at fit-to-screen.
Proof every legible label, sign and number. In the Topaz test, SeedVR2 turned a "2" into a "3" on a cable label, and neither model is safe on small text without a check.
If you’d like to order a framed print of one of my wildlife photographs, please visit the Prints page.
If you’d like to book a lesson or order an online photography course, please visit my Lessons and Courses pages.
