Resolution and Upscaling in 2026 (2K, 4K, Native vs Upscaled)

TL;DR: "AI image resolution" means something different on every model. Midjourney's HD mode tops out at 2048x2048 for a square image; Nano Banana Pro documents true 4K up to 6336x2688 at 21:9, but only through the API, AI Studio, or Vertex; FLUX.2 has no resolution tier at all, just width and height up to about 4MP; and Ideogram's own upscaler crops your image to its nearest supported ratio while enlarging it. Checked at each vendor's docs, September 2026.

Ask five different AI image tools for a 4K output in 2026 and you will get five different answers, and at least one of them will quietly hand you something smaller than 4K without saying so. AI image resolution is not one setting with one meaning across the field, it's five separate contracts, each with its own field name, its own casing rules, and its own idea of what "native" versus "upscaled" even covers. This piece checks each one against the vendor's own current documentation, not a six-month-old comparison chart, and separates what a model renders directly from what only comes out of a second, separate enlargement step.

What Do "Native" and "Upscaled" Actually Mean?

A native resolution is whatever pixel dimensions the model writes out in one generation pass, straight from your prompt. Set the parameter, get a file at that size, done. An upscale is a second, distinct operation applied to an image that already exists, sometimes built into the same tool and sometimes a different model or vendor entirely.

That distinction matters for three practical reasons. First, cost: Midjourney's own documentation prices an upscale in GPU minutes separately from the original generation, and states that upscaling can cost roughly twice as many GPU minutes as generating the initial image. Second, fidelity: an upscale is not guaranteed to be a clean pixel-for-pixel enlargement, several of the tools below either crop the image to a different aspect ratio or actively invent new detail while enlarging it, which a native render at the same pixel count would not do. Third, and most easily missed: of the five image models checked here, only Midjourney and Ideogram document a distinct upscale step at all. Nano Banana Pro, FLUX.2, and GPT Image 2 have no separate upscale endpoint in their published API references, resolution on those three is entirely a generation-time parameter, set once and rendered once.

What Does Midjourney V8.2's "2048²" Actually Mean?

It's the pixel dimension of one specific tier, at one specific aspect ratio, not a flat resolution Midjourney applies to every image. The current default Midjourney version is V8.2, which Midjourney's own Version article states became the default on July 24, 2026. V8.1, which shipped April 14, 2026, was itself the default version only from June 10 to July 23, before V8.2 replaced it, so a post citing "V8.1" as current is already two version bumps behind.

Inside V8.2 there are two separate resolution tiers, and the pixel numbers depend on which one you're in, not on a single flat "2048²" figure:

Source: Midjourney's Upscalers article, fetched September 3, 2026.

2048×2048 is real, but it is the 1:1 HD figure specifically, not a flat resolution Midjourney applies everywhere. Every other ratio scales differently, and Midjourney publishes exact pixel counts for exactly four ratios; anything you see for a fifth is someone's extrapolation, not a sourced number.

HD is not the default. You turn it on, in the settings panel or with --hd after your prompt, and it costs more: Midjourney's Version article states HD costs 1.3 minutes of GPU time, compared to SD which costs 0.8 minutes. SD is the version Midjourney's own docs currently describe as default, after HD briefly held that position in a since-reverted announcement earlier in 2026. Your own settings panel is the authoritative source for which one you're actually generating.

Two more limits worth knowing before you plan a shot list. Midjourney's own feature-compatibility chart lists the maximum aspect ratio as 14:1 across V6 through V8.2, with a parenthetical for HD specifically: 4:1 for HD. Push past that ratio in HD and you'll need to fall back to SD. And any inpainting or outpainting tool, Pan, Zoom Out, or Edit/Vary Region, downscales an HD image back to SD; Midjourney's docs say you then need to run the Upscale options again to get back to HD resolution. Editing an HD image is not resolution-neutral.

Does Nano Banana Pro Actually Output Native 4K?

Yes, for the specific model called Nano Banana Pro, and it's documented as a parameter rather than implied by marketing copy. But "Nano Banana" is Google's umbrella marketing name, and as of this writing it covers four distinct models with four different resolution ceilings, not one. Google's Gemini API image-generation guide lists them: Nano Banana 2 Lite (gemini-3.1-flash-lite-image, 1K only), Nano Banana 2 (gemini-3.1-flash-image, 0.5K through 4K), Nano Banana Pro (gemini-3-pro-image, 1K through 4K), and the legacy Nano Banana (gemini-2.5-flash-image, fixed output, no size control at all). Writing "Nano Banana" as though it names one resolution ceiling is the single most common mistake in circulation about this model family.

For Nano Banana Pro specifically, resolution is set through image_size inside response_format, and Google is strict about the exact string: "You must use an uppercase 'K' (e.g. 512px (05.K), 1K, 2K, 4K). Lowercase parameters (e.g., 1k) will be rejected." (Reproduced exactly as published, "05.K" is Google's own typo for 0.5K, appearing twice on the live page as of this fetch.) Send "image_size": "4k" in lowercase and the request errors, it does not silently fall back to a smaller tier.

4K is genuinely documented, with an exact pixel table per aspect ratio, not a marketing round number:

Source: Google's Gemini API image-generation guide, fetched September 3, 2026.

Notice the ceiling isn't 4096². 1:1 is 4096×4096, but both 16:9 and 21:9 produce a larger total pixel count than the square figure. Quoting "4096²" as Nano Banana Pro's maximum understates what wider ratios actually deliver.

Here's where the "real 4K, gated" part comes in. Even though gemini-3-pro-image genuinely renders 4K, the consumer Gemini app does not hand you that file. Google's own Gemini Apps help center states that images download at 2K resolution on a Google AI plan, or at 1K without one, and that cap applies regardless of which Nano Banana model drew the picture. The app's high-resolution preview is not the same as the download you get. To actually receive a 4K file, you need Google AI Studio, the Gemini API, or Vertex AI, and you need to set image_size explicitly yourself; the app's own "Redo with Pro" action does not document lifting that cap. Our 4K image workflow guide walks through the DPI math for turning these pixel counts into an actual print size once you have the file.

Does FLUX.2 Have a Resolution Tier or an Aspect-Ratio Field?

Neither. FLUX.2 handles resolution completely differently from either model above. Checked directly against the live FLUX.2 API schema, the request body for flux-2-pro and flux-2-max (Flux2Inputs) exposes exactly two size fields, width and height, each an integer with a documented minimum of 64px and no maximum stated in the schema itself. There is no aspect_ratio field, no resolution tier enum, and no negative_prompt field anywhere in the spec, a grep for "negative" across the whole file turns up exactly two unrelated hits about outpainting pixel offsets.

That's a real contrast with how BFL markets the same model: FLUX.2's own documentation page describes the family as offering "flexible aspect ratios", while the API schema backing that same page has no aspect_ratio parameter at all, only raw pixel dimensions you compute the ratio from yourself.

{
  "prompt": "A weathered leather journal on a walnut desk, warm window light, soft shadow, product photography",
  "width": 2048,
  "height": 2048
}

For a total-pixel ceiling, BFL's own docs describe the FLUX.2 model family as producing "up to 4MP" output, which a 2048×2048 square lands close to. There's no named "4K" toggle the way Midjourney or Nano Banana Pro expose one, you get there by choosing width and height yourself and staying under that ceiling.

The flex variant adds two parameters the standard tiers don't have: steps (an integer from 1 to 50, default 50) and guidance (1.5 to 10, default 5.0), both confirmed in the same schema. Neither controls resolution directly, but both affect how much detail the render actually contains at whatever pixel count you set, which matters if you're rendering close to the ceiling and expect the extra pixels to carry real information rather than padding.

Which Models Actually Resize Your Image Without Telling You?

Of the five checked here, only one documents it outright: Ideogram. GPT Image 2 uses a size field, either exact pixels or the string auto (the default). OpenAI's own image-generation guide lists 3840×2160 and 2160×3840 as its documented 4K landscape and portrait sizes, alongside a 2048×2048 "2K square" option. But the guide's size constraints are stricter than "just pick any pixel count": maximum edge length must be less than or equal to 3840px, both edges must be multiples of 16px, the long-to-short edge ratio must not exceed 3:1, and total pixels must fall between 655,360 and 8,294,400. Land outside any one of those and the request fails, there's no silent rounding documented on this page.

Ideogram behaves differently again, and says so plainly. Its consumer docs describe Aspect Ratio, Dimensions, or Resolution as three different controls depending on the model, workflow, and generation tier you're in, and state outright: "Ideogram may normalize your requested dimensions to a supported ratio, model size, or resolution tier." That's three separate normalizations named in one sentence, ratio, model size, and resolution tier, documented as real behavior rather than something users have to reverse-engineer from mismatched outputs.

Ideogram's separate Upscale feature (a paid-plan tool, distinct from generation) makes the same normalizing behavior even more concrete. Ideogram's own Upscale docs state plainly that upscale crops images to the nearest supported aspect ratio before enlarging, up to 2×, and walk through a worked example: a custom 17:12 image at 1088×768 gets cropped to the nearest supported ratio, 3:2, and comes out at 2304×1536. You don't get your original framing back at a larger size, you get the nearest ratio Ideogram supports, enlarged. Two sliders, Resemblance and Detail, then control how much the upscale is allowed to invent versus preserve, a low Resemblance value gives the model room to fix small defects rather than just add pixels.

Native Ceilings, Side by Side, Dated

Read that "highest pixel count" row carefully before you pick a winner: it's the largest figure each vendor publishes across the ratios it documents, not a like-for-like comparison of the same aspect ratio five times over. A model that documents 4K only at 21:9 is not directly comparable to one that documents its ceiling at 1:1.

When Does Upscaling Actually Beat Generating Native?

If a vendor offers both a larger native tier and an upscale path to the same pixel count, generating native is almost always the safer choice, you skip the extra cost Midjourney's own docs describe and you skip whatever cropping or reinterpretation an upscale step might apply. Reach for an upscale instead when one of three things is true: the image already exists and re-prompting from scratch would lose a composition you want to keep, you need to go past a model's documented ceiling entirely (none of the five models checked here publish a way to exceed their own native tier through the API itself), or you're deliberately using the upscale's own creative parameters, Ideogram's Resemblance and Detail sliders, or Midjourney's Creative upscaler, to fix a specific flaw rather than just add pixels.

What's notably absent across every image model checked here: a documented "go past native" upscale endpoint of the kind several AI video vendors publish (a separate resolution-boosting API call with its own pricing). For image generation in 2026, resolution is overwhelmingly a generation-time parameter, not a downstream service. If you need genuinely larger than any of these ceilings, a third-party image upscaler outside these five vendors is the honest next step, and none of the vendors above document one of their own that goes further than what's in the tables above.

A Prompting Checklist for Getting the Resolution You Actually Want

Before you send the prompt, confirm these four things against the model you're actually using, not against a comparison post from six months ago:

1. Identify the exact model, not the marketing family. "Nano Banana" alone tells you nothing, confirm whether you're on Lite, 2, or Pro, each has a different ceiling.
2. Set the resolution parameter explicitly, and match its exact casing. 4K and 4k are not interchangeable on Nano Banana Pro; --hd is a flag, not a number, on Midjourney; FLUX.2 wants raw width/height integers, not a ratio string.
3. Check whether your delivery surface caps the file below what the model generated. A native 4K render from Nano Banana Pro through the API is not the same file you'd get from the consumer app's download button.
4. Decide before you upscale whether cropping or added detail is acceptable. If the tool documents cropping to the nearest ratio (Ideogram) or documents that the upscaler adds new detail (Midjourney's Creative mode), that's a deliberate trade, not a bug, but only if you know it's happening.

Locking a seed alongside your resolution setting is the cleanest way to isolate what a resolution or aspect-ratio change actually did to a composition, without a second, unrelated variable moving at the same time; our seeds and reproducibility guide covers what each vendor actually promises there, and it is less than most tutorials imply. And if your workflow spans both images and video, resolution fields do not carry over between the two, our AI video resolution settings guide covers the video side, which uses entirely different parameter names and ceilings from everything in this piece.

None of this is settled once and left alone. Midjourney's Version article was itself updated within the last few days of this piece going out, and a resolution table that's accurate today can be one release note away from wrong. Check the parameter name, the exact casing, and the ceiling against the vendor's own current docs before you build a workflow that depends on any of the numbers above, Prompt Architects' free image prompt tools and FLUX.2 presets are built to the current parameter shapes, and get rebuilt the moment a vendor changes one.