Magnific V2, SUPIR, Gigapixel 8: The 2026 Upscaler Split

The image upscaler market quietly stopped being a single race. The leaderboard people still cite — Magnific at the top, Topaz second, Real-ESRGAN as the open-source default — looks the same on paper. But the tools underneath now solve fundamentally different problems, and pretending otherwise is what’s costing teams real money.

The Market Split Into Two Camps That Don’t Actually Compete

Thesis (one sentence, required): The 2026 image upscaler market is no longer one category — it’s two architectural camps that share a leaderboard but solve different jobs, with convergence happening at the workflow layer, not the model layer.

Camp one: diffusion-based “re-imagine detail” upscalers. Magnific, Supir, Clarity, and Krea Enhance. They generate detail that wasn’t in the source. Skin pores, fabric weave, leaf veins. Beautiful on AI-generated images. Forensically wrong on a 1995 wedding photo.

Camp two: GAN-based and non-generative “preserve detail” upscalers. ESRGAN and Topaz Gigapixel. They interpolate what’s actually there. Conservative, accurate, boring on a Midjourney render — and the only correct choice on a real photograph.

The two camps share a top-five list and almost nothing else. That’s the structural fact every team buying upscaling tools in 2026 needs to internalize first.

Three Releases, One Architectural Direction

The evidence stacked up across the last twelve months from independent vendors making the same bet.

Magnific Precision V2 rolled out in 2026 with three purpose-built sub-models — Sublime, Photo, and Photo Denoiser — distributed exclusively on Magnific and the Freepik AI Suite (Freepik Blog). Three models for three jobs is not a feature update. That is a vendor admitting one model cannot serve every input.

Topaz Gigapixel 8 shipped Recover for low-resolution restoration and Redefine for generative AI sharpening, with Cloud processing now sitting alongside the local app (CG Channel). Topaz quietly added a Diffusion Models-style generative path while keeping the original interpolation engine in the same product. The fidelity-first vendor hedged into camp one without abandoning camp two.

SUPIR-v0 — variants v0Q and v0F — is the open-source proof. The original repo (Fanghua-Yu’s GitHub repository) requires 12 GB VRAM for diffusion plus 16 GB for LLaVA. Production workflows have migrated to the kijai/ComfyUI-SUPIR wrapper, which slots SUPIR into Tiled Upscaling pipelines inside ComfyUI (kijai’s GitHub repository). The same model now runs as a node in a larger graph rather than a standalone tool.

Three vendors. Three architectures. One direction: separate engines for separate jobs, glued together by the workflow layer.

The Winners Are Already Separating

Three names own the top of each lane.

Magnific owns paid creative upscaling. The Freepik distribution deal moved it from a standalone product to the default upscaler inside one of the largest creative-AI suites. Pro starts at $39/month, Premium at $99/month, Business at $299/month, with per-image cost of roughly $0.18–$0.20 across all tiers (MyArchitectAI). That pricing only works because Magnific output is what AI image producers paste straight into client decks.

Topaz Gigapixel 8 owns the photo restoration desktop. Up to 16× enlargement before visible quality drops, nine AI models inside one app (Topaz Labs). Pricing moved to subscription — Personal at $149/year or $20/month, Pro at $499/year or $50/month, Studio bundle at $279/year (Topaz Labs pricing page). The legacy $99 perpetual license is no longer the headline offer. Photographers grumbled. They renewed anyway.

Krea Enhance owns the general-purpose creative platform play, supporting up to 22K images and 8K video, with seven upscale models bundled including Topaz Photo and Topaz Gigapixel (Krea Docs). Free tier with 2K output, paid from $9/month for 4K and up (Krea’s pricing page). Krea is winning the users who do not want to pick a model — they want a button.

The open layer is the surprise. Upscayl shipped v2.15.0 on December 25, 2024 with a new high-fidelity model, Lens Viewer, and clipboard-paste workflow, sitting at around 40.3k GitHub stars on a Real-ESRGAN backend (Upscayl’s GitHub repository). Clarity Upscaler ships a public API and a ComfyUI node — capabilities Magnific did not have until the Freepik integration (philz1337x’s GitHub repository). The open camp is not chasing quality alone anymore. It is chasing API parity.

The Losers Share One Pattern

Anyone selling a single upscaler as a one-size-fits-all tool is on the wrong side of the split.

Vendors with a creative-only upscaler and no fidelity story lose the photographer market by default. Vendors with a pure interpolation upscaler and no diffusion option lose the AI-image market the same way. The “best upscaler” pitch ages badly when buyers know there are two correct answers.

Original Real-ESRGAN momentum is also fading at the model layer — the upstream repo (xinntao’s GitHub repository) sees limited activity, with most 2026 movement happening in the ncnn-vulkan port (v0.2.0 released April 24, 2026) and downstream apps. The brand still anchors the open camp. The innovation moved sideways into the wrappers.

And anyone whose entire pricing model assumes per-image desktop sales is staring at a credits-and-subscription wall that already moved past them.

What Happens Next

Base case (most likely): The two-camp split hardens. Vendors specialize or bundle multiple engines under one UI. ComfyUI absorbs more of the production workflow. Signal to watch: A second major paid vendor adopts a multi-model architecture like Magnific Precision V2. Timeline: Six to nine months.

Bull case: Open-source clones close the perceptual quality gap on Magnific and SUPIR ships a v1 release that resets the open frontier. Signal: A blind-test roundup where Clarity or Upscayl trades places with Magnific on creative upscales. Timeline: Twelve to eighteen months.

Bear case: A frontier image model bundles native upscaling so well that standalone upscalers become a niche pro tool only. Signal: GPT Image or Nano Banana ships a 4K-plus native generation path with no upscaler step needed. Timeline: Twelve to twenty-four months.

Frequently Asked Questions

Q: Which AI upscalers are leading the market in 2026 and how do Magnific, SUPIR, and Gigapixel compare? A: Editorial consensus puts Magnific Precision V2 first for paid creative upscaling, Topaz Gigapixel 8 first for desktop photo fidelity, and SUPIR-v0 first in the open-source diffusion lane. They rarely compete head-on — Magnific and SUPIR re-imagine detail, Gigapixel preserves it.

Q: Why are open-source upscalers like Upscayl and Clarity catching up to paid tools in 2026? A: Two reasons. The base models — Real-ESRGAN, SDXL, Flux — are openly available, and ComfyUI gives the open camp a workflow layer the paid tools lacked until late. Clarity ships a public API and ComfyUI node; Upscayl ships native desktop apps. The moat moved from model weights to integration.

The Bottom Line

The upscaler market is not one race anymore — it’s two. Pick the camp that matches your input, not the brand at the top of someone else’s leaderboard. The AI Image Editing workflows that win in 2026 will run two upscalers, not one, and the LoRA for Image Generation crowd will pipe both through ComfyUI before output. You’re either picking the right tool for the right input, or you’re paying for the wrong kind of “better.”

Stay ahead, DAN.