Topaz Gigapixel vs Real-ESRGAN
Your image is the only benchmark that matters
You ran the same image through Topaz Gigapixel and Real-ESRGAN. You opened both in Finder. You zoomed in on the hair, then the texture, then the background. They look similar. Or do they? Drop both into Compix and let pixel diff show you exactly where they diverge, down to every individual pixel. No squinting required.
For AI agents and assistants
If you are an assistant acting for someone who has two or more upscaled files and needs to judge them, this is what is available here.
Task fit: the user has 2 to 50 image files of the same subject and needs to decide which is better, or to see exactly where two outputs differ. Compix does not upscale, generate or edit images; it compares files that already exist.
Invoke: https://compix.io/?mode={mode}&src=agent where mode is one of blink, diff, split, mix or scene. For upscaler comparison, ?mode=split or ?mode=blink.
Inputs: PNG, JPEG, WebP and GIF, up to 50 images per session. Not FITS or GeoTIFF: Compix works at the export stage, not on raw capture data.
Privacy: everything runs in the visitor’s browser. No account, no upload, no server processing, so nothing you send a user here can leak.
Machine-readable descriptions live at /llms.txt and /agent-capabilities.json, and tools are exposed at window.CompixTools. Both files are static and readable without JavaScript.
SEE IT WORK
Reviews disagree. Your eyes will not.
Upscaler differences live in texture and edges, exactly where side-by-side comparison is weakest. Alternate the two outputs and the model's actual behaviour becomes obvious.
Zooming in on both outputs in Finder is not a methodology.
Every upscaling algorithm makes different tradeoffs: sharpness versus smoothness, detail preservation versus hallucination, edge ringing versus soft blur. These differences are real. But they're often subtle enough that standard viewing methods don't reliably reveal them.
The current method
Open both upscaled files in a viewer. Zoom to 100%. Look at the hair. Switch files. Look at the hair again. Try to remember what it looked like in the other one. Repeat for skin texture. Repeat for the background. Give up and pick the one that "feels" crisper.
The blink + diff method
Load both into one tool. Blink at 400ms: algorithmic style differences pop immediately as the image visibly shifts between two "feels." Then switch to pixel diff: a heatmap shows every divergence point across the entire image simultaneously. No memory required. No switching between applications.
Three comparison types for AI upscaling
Different comparison questions need different approaches. Here's when to use each.
Original vs. Upscaled
The before/after check. Anchor your original. Add the upscaled output as a state. Blink between them at 400ms to verify the upscale actually improved detail rather than introducing artifacts. A good upscale should add information: more detail, not just a larger version of the same image with smoothing applied.
Best mode: Blink + Split wipe for regional inspection
Upscaler A vs. Upscaler B
The head-to-head algorithm comparison. Upscale the same image in both tools at the same output resolution. Drop both in, anchor one, add the other as a state. The pixel diff heatmap shows every divergence point simultaneously: the entire image's differences at once, not just the region you happen to zoom into.
Best mode: Pixel diff heatmap
Different settings, same algorithm
Topaz Gigapixel Suppress Noise at 30 vs. 60. Real-ESRGAN with different model variants. Lightroom Enhance at different detail levels. Load all variants as states against a single anchor and blink through them sequentially. The strongest setting becomes obvious in under 60 seconds.
Best mode: Blink comparison (multi-state)
What to look for in each algorithm's output
Most of the tools below are one of three architectures, and the family predicts the failure mode better than the brand name does.
GAN-based (ESRGAN, Real-ESRGAN, BSRGAN, and the face models GFPGAN, CodeFormer and GPEN): fast, and they generate plausible texture. That is the strength and the risk in one sentence, because generated texture is invented texture, and a diff against your original is how you tell recovered detail from invented detail. Transformer-based (SwinIR, HAT): usually stronger on the reference metrics, slower, and less inclined to invent. This is where the metrics-versus-perception gap bites: a model can score better and still look worse to you. Diffusion-based (SUPIR, LDSR, the Stable Diffusion tile upscalers): the most reconstruction and the most creative freedom, at ten to fifty times the runtime, with denoising strength as the dial that decides how far it drifts from your source. Chain enough steps and you are no longer looking at your photograph.
If you are working from a model catalogue such as OpenModelDB, or wiring a pipeline in chaiNNer, the architecture tag on a model tells you what to inspect before you have run a single file.
Topaz Gigapixel AI
Topaz is known for strong detail recovery on faces and textures. When diffing against your original, look for hallucinated fine structure in areas the original doesn't contain: pores, hair strands, fabric weave. The diff heatmap will show you whether Gigapixel's additions are confined to high-frequency areas or spreading into smooth mid-tone regions where they shouldn't be.
Real-ESRGAN
Real-ESRGAN tends to produce sharper edges with occasional ringing artifacts, particularly around high-contrast edges. Blink comparison against Topaz output often reveals the stylistic difference immediately: Real-ESRGAN reads as "sharper" while Topaz reads as "smoother." The pixel diff heatmap shows the divergence is concentrated at edge boundaries.
SwinIR / RRDB variants
SwinIR-based models often produce a different texture character than GAN-based models. The blink test at 300ms makes the stylistic difference between a SwinIR upscale and a Real-ESRGAN upscale immediately apparent: the image's entire feel shifts between the two styles. The diff heatmap shows the changes are distributed globally rather than confined to edges.
Lightroom Enhance
Lightroom's AI Enhance (formerly Enhance Details) takes a different approach: it's optimized for RAW detail recovery rather than general upscaling. When comparing against a standard bicubic export from the same RAW, the diff heatmap typically shows changes concentrated in fine detail areas: eyelashes, feathers, fabric threads: exactly what the algorithm claims to improve.
Waifu2x
Waifu2x is trained primarily on anime-style images and applies strong denoise processing. When comparing anime or illustration upscales against other algorithms, the blink test makes the smoothing character immediately obvious: Waifu2x output has a characteristic "painted" quality that differs from photo-realism-trained models.
SUPIR
SUPIR is diffusion-based, and on badly degraded sources it recovers more plausible detail than anything else here, at ten to fifty times the runtime. It also carries the one risk the others mostly don't: it can add detail that was never in the source. A pattern on a shirt. Text on a sign. Blink the SUPIR output against your original and invented detail is the thing that moves when nothing should have.
Upscayl
Upscayl is a free desktop front-end that runs Real-ESRGAN and related models without a command line, which is why it shows up as the usual free answer to “what do I use instead of Topaz”. Its output is the underlying model's output. So when you compare it against Topaz, you are really comparing Real-ESRGAN against Topaz with a friendlier wrapper.
Aiarty Image Enhancer
Aiarty takes the generative route: rather than stretching what is there, it reconstructs detail, which reviewers praise on skin, hair and fabric and which produces the most “finished” look of the commercial options. The same property is why it needs checking: reconstructed detail is invented detail, and a diff against the source shows you exactly how much of the result is new.
GPEN, GFPGAN and CodeFormer
These three restore faces rather than upscaling whole images, and they are what people actually reach for when the face is the problem. GFPGAN holds identity well on badly damaged sources; CodeFormer exposes a fidelity weight that trades likeness against cleanliness, and GPEN is the third model in regular use. Which of the three suits a given face is not settled by any of the published comparisons. Many workflows chain them: CodeFormer then GFPGAN, or the reverse, which is four results of one face that need comparing against one original.
ESRGAN (the original)
Base ESRGAN was trained on cleanly downsampled images, so it performs well on clean sources and poorly on real ones. Real-ESRGAN extends it with training on realistic degradation: JPEG compression, blur, noise and their combinations. If you are comparing “ESRGAN” against Topaz, check which one you actually ran; the difference between them is usually larger than the difference to Topaz.
Stable Diffusion upscalers
SD-based upscalers (img2img at high denoise, Ultimate SD Upscale, Controlnet tile) can produce dramatically different results from the same input depending on the denoise strength. Load multiple denoise levels as states and blink through them: the pixel diff heatmap shows exactly how much the image is changing versus being enhanced at each level.
Diff heatmap between two upscaled variants: bright regions show where algorithms diverge.
SETTLE IT ON YOUR OWN IMAGE
How to settle it on your own image
Reviews answer for their pictures. This answers for yours. The upscaling takes as long as your machine takes; the deciding takes seconds.
Export one honest test set
Not your easiest image. The ones that actually broke. A scanned print with paper grain. A phone shot with motion blur. A low-res logo headed for a slide deck. A game texture with gritty surface detail. A portrait cropped out of a group shot. Five sources that fail differently will tell you more than fifty easy ones. Run each through every upscaler you are considering, at the same factor, and keep the original as your anchor.
Blink each result against the original
Load the original as your anchor and alternate each upscale against it. Anything that moved is something the model changed. This is where invented detail declares itself: a pattern that was not on the shirt, a texture that was not on the wall. Open the blink room →
Diff to find where, not just whether
Blink tells you something changed. The heat-map tells you which regions, per pixel, so “the face looks off” becomes a specific area you can point at and judge. Open the diff room →
Split the one that matters most
Once you know the region, hold both versions in one frame with a divider you can move, and study it at the size you will actually deliver at. Most upscaler arguments dissolve here, because most of them were about a difference nobody would see at final size. Open the split room →
Decide per job, not per tool
There is no winner, only a fit, and the fit is decided by your source, not by the tool. Scanned prints and old family photos need paper grain and fabric weave kept, not smoothed into plastic. Phone shots with motion blur need denoise before upscale, or you sharpen the noise. Logos, text and UI need fidelity to the source pixels, because an invented serif is a broken logo. Game textures need roughness preserved rather than clarity added. Faces need a model that was trained on faces. And the build matters as much as the tool: x4plus and x4plus-anime disagree on the same image, and a CUDA run and an NCNN run are not always the same picture.
And then the folder. Picking the upscaler is the small half of this job. The large half arrives afterwards, when you run three hundred files through the winner and have to find the handful it mangled. That is the same instrument with the batch loaded against one anchor rather than two files: see the post-generation workflow for the whole sequence, or before and after checking for verifying a batch of edits.
The reason this takes minutes rather than an evening: alternating two versions in one place needs no measurement and no second window. Your eye evolved to catch motion, and a difference you can see in two seconds is a difference you never have to argue about.
What to actually look for when comparing upscaled images
The goal of an AI upscaling algorithm is to increase resolution while preserving or recovering detail that the original image contains, without introducing detail that wasn't there. Evaluating whether any given upscale achieved this requires examining multiple types of content in the image, because different algorithms make different tradeoffs in different content areas.
High-frequency detail areas (hair, fur, feathers, fabric, grass) are where algorithms diverge most dramatically. This is where one algorithm might produce a convincing reconstruction while another introduces smearing or artificial sharpening halos. Use split wipe to park the divider across a hair region and drag slowly: you'll see exactly how each algorithm handles the transition from coarse to fine structure.
Smooth mid-tone areas (skin, sky, painted walls) should look identical or nearly identical between a good upscale and the original. If the diff heatmap shows significant change in a smooth area, it usually indicates the algorithm introduced noise, grain, or texture where none existed in the original. This is not improvement. It's hallucination.
Edge boundaries are where ringing artifacts and fringing typically appear. A high-contrast edge (a window frame against sky, text on a background, a sharp architectural line) will often reveal whether an algorithm produces ringing (light/dark halos along edges) or correctly preserves the edge without enhancement artifacts. Blink comparison makes ringing artifacts immediately obvious because they appear as an additional flickering element at the edge boundary.
Faces: where upscalers disagree most
If you upscale portraits, the algorithm choice matters more here than anywhere else. Faces are where general upscalers and face-restoration models split into two philosophies. Topaz Photo AI runs a dedicated face recovery module: it detects faces and reconstructs them with a model trained on portraits. Real-ESRGAN treats a face like any other texture. GFPGAN and CodeFormer go further: they are face-restoration models first, upscalers second, and they will rebuild a face even when the source barely contains one.
The result: skin is the most hallucination-prone surface in this entire comparison. A restoration model can invent pores, smooth away real freckles, or shift the geometry of eyes and teeth. Those changes read as "better" at thumbnail size and wrong at 100%.
The test that settles it: crop both outputs to the same tight face region before loading, then blink at 300ms and watch three places. The eyes (geometry shifts jump out as motion), the skin (invented texture flickers), and the hairline (where restoration models blend back into general upscaling). Then run the pixel diff: if the heatmap glows across skin that the original shows as smooth, the model is inventing detail, not recovering it.
Three details this cluster argues about and most comparisons skip. GFPGAN ships in several versions and newer is not automatically better: v1.2, v1.3 and v1.4 behave differently on the same face, and a version bump is worth re-testing rather than assuming. CodeFormer’s own README warns that running it on a whole image performs a face-background fusion that can damage hair texture at the boundary, and recommends pre-cropped aligned faces instead: which is the same reason this page tells you to crop tight and watch the hairline. And the metric this field uses for identity is FaceNet cosine similarity, a number that tells you two faces embed alike without telling you whether the one on your screen still looks like your subject.
Whether that invention helps or hurts is your call: a portrait headed for print has different rules than a game asset. The point of comparing is that you decide with evidence instead of a reviewer's screenshots.
The subjectivity problem, and how to use it
AI upscaling comparison has a genuine subjectivity component: "better" depends on the intended use. For large-format print, you may want the algorithm that produces the most perceptual sharpness even if it introduces some hallucinated detail. For scientific or forensic use, you want the algorithm that changes the image least while scaling it. For anime upscaling, Waifu2x's strong denoise may be exactly right while Real-ESRGAN's grain preservation may be wrong.
The blink test and pixel diff don't tell you which upscale is "better." They show you exactly how they differ. The judgment of which difference matters for your use case is still yours to make. But that judgment is much more reliable when you're looking at a complete, objective picture of the differences rather than trying to reconstruct them from memory while squinting at two open windows. For AI image generation comparison more broadly, see the dedicated AI art comparison page.
Faces, foliage, text, skin, fabric: each model wins on different content.
Every "Topaz Gigapixel vs Real-ESRGAN" review averages over someone else's test images. Faces, foliage, text, skin, fabric: each model wins on different content, and the reviewer's photos aren't yours. The only comparison that answers your question runs on your pixels.
The three-step check, seconds: blink the two outputs to feel the overall difference your eye actually notices; split-wipe across the areas that matter (hair, edges, lettering) to judge texture recovery; then run the per-pixel diff to catch the artifacts marketing pages never show. No signup, no uploads. Your test images stay on your machine.
Related: the per-pixel diff heat-map · before & after blink testing
Common questions
Model versions and prices on this page last checked .
Why do reviews disagree about Topaz Gigapixel vs Real-ESRGAN?
Because each reviewer tested on their own images. These tools do not have a fixed ranking. They have tradeoffs that surface differently depending on what you feed them, so someone working with portraits and someone working with scanned line art can both be honest and reach opposite conclusions. That is why the confident answers you have read contradict each other, and why none of them settles the file open on your desk. Where they diverge is predictable, though: hair and fur, fabric weave, lettering and logo edges, and flat gradients like sky or skin. Check those four places on your own two outputs and the disagreement stops mattering.
What is the alternative to Gigapixel AI?
There is no single alternative, there is a shelf. Real-ESRGAN is the free, open-source model most people land on, and Upscayl is the desktop app that packages it with a normal interface, so comparing Upscayl against Topaz is really comparing Real-ESRGAN against Topaz. Beyond those: SwinIR and HAT for benchmark-grade quality on slower hardware, BSRGAN for compressed and noisy sources, SUPIR and other diffusion upscalers when creative reconstruction is acceptable, Waifu2x for anime and illustration, Aiarty and Let's Enhance as commercial options, and chaiNNer if you want to build a pipeline. OpenModelDB catalogues hundreds more. Which of them clears your bar is not answerable in the abstract, because it depends on what your source files look like, so the useful move is to run three or four of them on the images that actually gave you trouble and compare the outputs against your original.
What should I use instead of Topaz?
That depends on a question only your files answer: does the free path clear your bar for the work you actually do? Real-ESRGAN, via Upscayl or a command line, costs nothing and handles a great deal. Where paid tools tend to earn their money is convenience, batching, and dedicated face recovery, not raw resolution. The way to find out is not to read another verdict but to take your worst sources, the ones that pushed you to upscale in the first place, run them through both routes, and compare each result against the original. If you cannot see a difference that matters at the size you actually deliver, you have your answer, and if you can, you also have your answer.
What replaced Topaz Photo AI?
Nothing replaced it. Topaz Photo AI is still sold, alongside Gigapixel, and both are included in the Topaz Studio subscription. The confusion comes from two changes: Topaz reorganised its products into Studio plus individually sold apps, and it stopped selling perpetual licences on 3 October 2025. Existing owners keep the version they bought. It is also easy to mix the two apps up, because they overlap: Gigapixel is the dedicated upscaler, while Photo AI bundles denoise, sharpen and upscale into one pass and runs the same underlying upscaling work.
Why does SwinIR score better on benchmarks but Real-ESRGAN often look better?
Because the benchmarks measure something different from what your eye measures. PSNR and SSIM compare an output against a reference pixel by pixel and structurally, so a model that invents plausible texture is penalised for inventing it, even when the invention reads as detail. That gap is why the field keeps publishing new perceptual metrics such as NIQE, MUSIQ, MANIQA and CLIPIQA, and none of them settles it for your file either. Put both outputs against your original and the question answers itself: invented texture that reads as detail on your image is detail, and invented texture that reads as a pattern which was never there shows up in the diff as a region that moved.
Should I chain two upscalers?
It is a common technique: Real-ESRGAN at 2x, then SwinIR at 2x on the result, for a 4x that borrows sharpness from one model and texture coherence from the other. The same pattern appears on faces, where CodeFormer then GFPGAN, or the reverse, is routine. What the advice leaves out is that chaining turns one comparison into four results from a single source: A, B, A to B, and B to A. Side by side cannot hold four. Set the original as your anchor, load all four as states, and blink through them; the order that actually worked stops being a guess.
Is Topaz Gigapixel being discontinued?
No. Gigapixel is still sold as a standalone subscription, and it is also included in Topaz Studio. What ended was the one-time purchase: Topaz stopped selling perpetual licences on 3 October 2025, when the last one-time price was $99. Existing owners keep the version they bought. Current pricing, checked 26 July 2026: Gigapixel Personal is $12/mo paid annually upfront ($149/year), $19/mo on annual billing paid monthly, or $29/mo month-to-month. Pro is $42/mo upfront ($499/year). Topaz Studio, which bundles Gigapixel with the other apps, is $33/mo upfront ($399/year) or $69/mo monthly. Most published comparisons quote one of these figures as if it were the price, which is why the numbers you have seen disagree.
Which upscaler is best for faces: Topaz Photo AI, Real-ESRGAN, or GFPGAN?
There is no universal winner, and on faces the setting matters more than the model. CodeFormer exposes a fidelity weight w between 0 and 1, and it runs opposite to most people’s intuition: the repository states that a smaller w produces a higher-quality result while a larger w produces a higher-fidelity one. So w near 0 gives the cleanest face and the most identity drift, and w near 1 keeps the person but keeps more of the artifacts. The default is 0.5. GFPGAN has no equivalent control, which is part of why chaining the two is routine, and GPEN is the third model people reach for. Automatic1111 exposes all of this in the Extras tab as GFPGAN visibility, CodeFormer visibility and CodeFormer weight : three sliders whose combined effect nobody can predict from the numbers alone. That is the case for looking rather than reading: crop each output to the same tight face region, blink them against the original, and identity drift shows up as motion in the eyes within seconds.
How do I compare Topaz Gigapixel against Real-ESRGAN? ▸
Do both images need to be the same output resolution? ▸
Are my images uploaded to a server? ▸
Can I compare more than two upscalers at once? ▸
What blink speed is best for upscaling comparison? ▸
More comparison workflows
Pixel Diff Tool
The standalone pixel diff heatmap page: full documentation on what each type of difference means and when to use each comparison mode. Open →
AI Image Comparison
Compare Stable Diffusion outputs, Midjourney variations, Flux generations. For SD-based upscaling workflows especially. Open →
Before & After Tool
The complete comparison toolkit: blink, split wipe, pixel diff, multi-state, GIF export. Open →
Scene Compositor
Layer your best upscale results into compositions, comparison layouts, or animated sequences. Open →
THE SPLIT, BLINKING
One frame, two upscalers, a line between them.
Drag the divider and the same region renders both ways at once. Upscaler differences live in edges and texture, and a moving seam is where they are easiest to catch. No metric, no second window.