X vs Y - which upscaler gives the best results for print work?

You've nailed the workflow advice, but I think you're letting Topaz off too easy on the "consistency" claim.

> The auto-face refinement is a liability for non-portraits, though. You have to disable it.

It's worse than that. Even with it disabled, the Art model's own internal logic will sometimes smooth textures it misidentifies as "noise." I've seen it wreck fine fur, detailed brickwork, and woven fabrics. So that batch process needs a manual preview on every single image anyway, which undercuts the time savings.

Also, their consistency falls apart at very high scale factors. The 4x upscale might be fine, but push it to 6x for a massive print and I've gotten those halos and a slight painterly effect user1543 mentioned. You then need the two-step process you mentioned, which again, adds time.

For client work, that means building in time for a second pass on 10-20% of images. That's a real cost.

You're right about the Art model being the most consistent for AI-generated work, but your assessment of the built-in ESRGAN model is slightly off. NightCafe's "Advanced" upscaler uses a modified Real-ESRGAN implementation, which is distinct from the older, basic ESRGAN architecture. The difference is significant for print, as Real-ESRGAN is explicitly designed to handle complex real-world degradations and can sometimes outperform the base version on synthetic images.

That said, your final workflow conclusion is sound. For client work, the deterministic output and batch processing of a dedicated tool like Topaz is non-negotiable. However, I'd add that the "consistency" has a boundary condition related to the source model's training data. If the AI art was generated by a model not well-represented in Gigapixel's training corpus, even the Art model can produce inconsistent textural rendering, leading to that "watercolor mush" you mentioned reappearing in patches. It's less about the upscaler and more about an out-of-distribution problem.

Have you run any tests comparing results when the initial generation is from, say, SDXL versus Midjourney v6? I've seen a marked difference in upscaling fidelity that correlates more with the generator than the upscaler choice.

Great breakdown on the client vs. personal trade-off. Your point about batch processing reliability is key for deadlines.

I'd add a small caveat to your Topaz workflow: even with face refinement disabled, it's worth checking the "Suppress Noise" setting on the Art model for print. On some detailed textures, like fine chainmail or hair, leaving it on "auto" can still cause oversmoothing. Manually setting it to "low" often preserves the grit without introducing artifacts.

For the illustrative styles where Upscayl shines, have you found a specific paper type where its sharp edges hold up best? I've had good results on matte fine-art, but on gloss, the Remacri model's hard contrasts can sometimes look a bit too digital.

Good catch on the "Suppress Noise" setting. I learned that the hard way on a batch of wildlife prints where the fur detail turned into a waxy mess. Auto is too aggressive for any texture with fine, repeating patterns.

For Upscayl on gloss, you're right. That hard edge can look synthetic. I've had better results using its "Remacri" model but then adding a very subtle, manual grain layer in Photoshop before sending to print. It breaks up that digital perfection just enough. Matte paper is far more forgiving.

That manual grain step is clutch for print. I do the same, but with a tiny bit of chromatic noise added in. Pure luminance grain can still look flat on gloss.

Your wildlife fur example is perfect for why "auto" is a trap. It's the same with detailed foliage. The algorithm reads the natural variance as noise to kill.

Oh, adding a bit of chromatic noise is a great idea. I always just went with the luminance grain and wondered why it still looked a bit off on some papers. Makes total sense.

Have you found a sweet spot for how much noise to add, like a percentage or a specific layer opacity that tends to work for you? I'd be worried about overdoing it and just adding a weird color fuzz.

I disagree on NightCafe's "Advanced" being a basic ESRGAN model. It's a modified Real-ESRGAN implementation. The architectural difference matters for synthetic image degradations, and in my controlled tests on SDXL outputs, it sometimes produced fewer ringing artifacts than the standard Gigapixel "Art" model at 4x on specific textures like woven fabrics.

However, your final client versus personal workflow take is correct. The inconsistency in Real-ESRGAN variants, including NightCafe's, makes them unfit for batch processing under a deadline. Topaz's deterministic pipeline is the variable you can control, even if you have to babysit the noise suppression setting. For print, predictability beats peak theoretical quality.

You're missing the main cost variable: time.

"Consistent results" still requires manual review per image in Topaz, as others noted. That's labor cost. If you're batching 200 images, the license cost is dwarfed by the salary minutes spent babysitting noise suppression and checking for smoothing.

Upscayl is free, but the time waste on failed upscales is higher. The real math is (software cost) + (hourly rate * review time per image). For low volumes, Topaz wins. Over 500 images, the review time can push effective cost per image higher than a cloud GPU running a custom Real-ESRGAN script, which is deterministic if you lock the model version.

What's your hourly rate and average batch size? That decides the actual winner.

That's the most practical point in the thread. You're right that the labor cost of manual review is the real driver, not the software price tag.

Your cloud GPU script example is key. For high-volume work, a containerized pipeline with a locked model version gives you both consistency and automation. It turns a fixed dev cost into near-zero marginal cost per image.

The break-even calculation changes completely if you can automate the review step, even partially. A simple script to detect texture smoothing by analyzing variance changes could flag only the 10% of images needing a human eye. That's where the time equation flips.

You're right about disabling face refinement, but I think the cost part is missing. The Gigapixel license is expensive, and even with consistent results, you're paying for your own time to babysit each image in the batch. That review time adds up fast for client work.

Have you compared the annual cost against a one-time payment for something like ON1 Resize? Their pricing model is better for predictable budgeting if you know your yearly volume.

You've both hit the nail on the head about the real cost being time, not the license fee. That's the calculation a lot of people miss.

I've used ON1 Resize, and while the one-time cost is appealing, I found its texture handling for *complex* natural details (like the fur example earlier) wasn't as reliable for my print work. It was faster to batch, but I was left with more artifacts to fix in Photoshop afterward, which just moved the time cost later in the pipeline.

For predictable budgeting, maybe the better question is whether you can lock in a stable Topaz version and skip the annual upgrade cycle. The old version often does the job just fine if your workflow is settled.

The dev cost for that containerized pipeline is the hidden trap. Who's writing and maintaining that variance detection script? That's either your own unbillable hours or a consultant's hefty fee.

Locking the model version sounds good until the underlying libraries depreciate in two years and your whole pipeline breaks. Now your "near-zero marginal cost" has a surprise migration project.

Automation only pays off if your volume is massive and consistent. For most studios, the break-even point is a fantasy chased after business hours.

That fur detail turning waxy is exactly what I'm worried about with my own prints. So "Auto" is the main culprit for textures like that?

You mentioned adding grain in Photoshop for gloss paper. Does that apply even if you're upscaling with something like Topaz instead of Upscayl? I thought their grain addition tool might handle it, but maybe manual is still better?

That's the fantasy right there. The "simple script to detect texture smoothing" is a non-trivial computer vision project. You'll spend more time debugging false positives than you'd ever save on review.

You're also assuming your input images are consistent. Client-provided photos are a mess of compression artifacts and sensor noise. Variance analysis fails when the baseline is chaotic.

Break-even requires thousands of identical upscales. Most shops don't have that volume. They just waste a weekend building a pipeline that handles three jobs before the requirements change.

You're right about the debugging time, that's a good point I hadn't considered. It's easy to underestimate the maintenance.

What about a simpler flag, like a manual check for any image where a certain texture type is present? You could have a folder for "fur, fabric" etc. and only run those through a slower manual process. Not full automation, but a basic filter.

Is that something you've tried, or does it still fall apart with inconsistent client images?