Image Size and Megapixel Settings by Model (2026)

TL;DR: Every image model caps size differently, and the control is never a word in your prompt. Midjourney defaults to 1024×1024 and needs --hd for 2048. FLUX.2 takes raw width/height up to 4 megapixels. Ideogram 4.0 traded 69 resolution presets for 38, but doubled its ceiling. Nano Banana Pro reaches 4096×4096. Below is the sourced table, per-model detail, and copy-paste settings for each. Checked September 2, 2026.

Why do image models cap resolution instead of scaling freely?

Because a diffusion or transformer image model is trained at a handful of specific pixel counts, and asking it to render far outside those counts is asking it to extrapolate past what it ever saw during training.

Black Forest Labs is the one vendor that puts a number on this in its own documentation, and it is scoped to a single tool rather than the whole FLUX family. Its erase-v1 model card states: "The model was trained on images at ~1 megapixel across 9 aspect ratios from 1:2 to 2:1." It adds that inputs close to these resolutions "produce the best results" while "significant deviations may reduce quality." No other vendor in this piece publishes an equivalent sentence, but the pattern behind every ceiling below is the same one BFL states out loud: a model has a native operating range, and a resolution or aspect-ratio field exists to keep your request inside it.

That is also why "8k, ultra high resolution" typed into a prompt does nothing on any surface checked here. Size is a request field, not a word choice. It sits in the prompt next to your subject and carries no more weight than "moody" or "cinematic" — covered in more depth in why an image comes out the wrong size.

The size and megapixel ceiling, by model

Every figure below comes from the vendor's own API specification or help documentation, fetched directly rather than summarized, with the access date noted per row.

What size does Midjourney actually generate at?

1024×1024 pixels at the default 1:1 ratio, standard definition. Midjourney's Image Size & Resolution article states that "Midjourney version 8.2 creates HD images that are 2048 x 2048 pixels (px) and SD images that are 1024 x 1024px." HD is not the starting point; the Version article describes turning it on "in the Version section of the settings panel on web, or by using the --sd and --hd parameters."

Midjourney publishes exact pixel dimensions for exactly four ratios, doubling cleanly from SD to HD: 1:1 is 1024×1024 → 2048×2048, 4:3 is 1232×928 → 2464×1856, 2:3 is 896×1344 → 1792×2688, and 16:9 is 1456×816 → 2912×1632. Every other ratio you have seen quoted was calculated by someone, not published by Midjourney.

The maximum aspect ratio also depends on which resolution you picked: 14:1 in SD, dropping to 4:1 the moment HD is on, per the Version article's Feature Compatibility chart. And --ar takes whole numbers only — Midjourney's own fix for a decimal ratio is to write --ar 139:100 instead of --ar 1.39:1.

For the full ratio table, the HD-default contradiction in Midjourney's own changelog, and what happens with unsupported ratios, the --ar parameter reference covers it in depth. GPU cost differences between SD and HD generations are covered separately in Midjourney's quality and speed modes.

What resolution does Nano Banana Pro output?

Up to 4096×4096 at a 1:1 ratio, on the 4K tier. Google's Gemini API reference documents an ImageConfig object with exactly two fields: aspectRatio (14 values, including 1:1, 16:9 and four extreme ratios up to 8:1) and imageSize (512, 1K, 2K or 4K, default 1K). That is the entire object — no steps, no guidance, no seed.

Nano Banana Pro is the branded name for gemini-3-pro-image, and it is not the only model reaching 4K. Google's resolution table describes gemini-3.1-flash-image (Nano Banana 2) generating "state-of-the-art 4K" output too; only gemini-3.1-flash-lite-image (Nano Banana 2 Lite) is 1K-only. Worth flagging, not resolving: the table's own column header reads "3.1 Pro Image", while the model's branding and model ID both say "Gemini 3 Pro Image" and gemini-3-pro-image, with no ".1" anywhere. Both are live on the same page.

A second surface exists for the same two controls: Google's newer Interactions API takes response_format: { aspect_ratio, image_size }, snake_case, instead of ImageConfig's camelCase aspectRatio/imageSize. Same settings, two naming conventions, depending on which endpoint you call.

The full resolution table across all ten aspect ratios, the pricing-per-tier breakdown, and the uppercase-K requirement are covered in more depth in how to prompt Nano Banana Pro and the Nano Banana Pro prompt generator.

How big can a FLUX image get?

Up to 4 megapixels on FLUX.2, and Black Forest Labs' own API specification, api.bfl.ai/openapi.json, shows exactly how that is enforced: the Flux2Inputs schema (covering flux-2-pro, flux-2-max, flux-2-flex and the klein variants) takes width and height as plain integers, minimum 64px each, with no aspect_ratio field at all. BFL's own documentation states output "up to 4MP" across the FLUX.2 family, and separately defines one megapixel as "1,048,576 pixels (1024 x 1024)" for pricing purposes, not a decimal million.

Older FLUX.1 [dev] and FLUX1.1 [pro] work differently. Their schemas cap width and height at 1,440px each, with a floor of 256px, and the field description states plainly: "Must be a multiple of 32." Defaults are 1024×768. FLUX1.1 [pro] Ultra breaks from both patterns again, taking an aspect_ratio string and reaching 4MP the same way FLUX.2 does.

Full presets, the 32-token generation budget, and why "8k" typed into a FLUX prompt does nothing are covered in the FLUX.2 prompt generator. The separate guidance parameter, which controls prompt adherence rather than size and is not exposed on every FLUX.2 endpoint, is covered in FLUX guidance scale explained.

How many sizes does Ideogram support, and how did that change?

Fewer presets, but a higher ceiling. Ideogram 3.0's generate endpoint documents a resolution enum, ResolutionV3, with 69 fixed values, the largest being 1024×1024, about 1.05 megapixels. Ideogram 4.0's ResolutionV4 enum has only 38 values, but tops out at 2048×2048, about 4.19 megapixels — nearly four times the ceiling from just over half the preset count. Both counts come directly from developer.ideogram.ai's published OpenAPI specification.

The two versions also disagree on how you request a shape at all. Version 3.0 accepts aspect_ratio (an enum like 16x9 or 1x1) as an alternative to resolution, and its own schema description states: "Cannot be used in conjunction with resolution." Version 4.0 dropped aspect_ratio entirely — the field does not exist on the /v1/ideogram-v4/generate request — leaving resolution as the only sizing control.

That size change rides alongside a parameter change that has nothing to do with dimensions: 4.0 also dropped seed and negative_prompt from the request, even though its response object still requires a seed field. If you also work with typography or non-Latin scripts on Ideogram 4.0, typography prompting with Ideogram 4.0 covers that same version gap from the text-rendering side.

What sizes and formats does GPT Image 2 support?

Any resolution that satisfies four published constraints, rather than a fixed list. OpenAI's image generation guide states them directly: the maximum edge must be "less than or equal to 3840px", both edges "must be multiples of 16px", the long-edge-to-short-edge ratio "must not exceed 3:1", and "Total pixels must be at least 655,360 and no more than 8,294,400". The guide also flags a soft ceiling inside that hard one: anything past 2,560×1,440 (3,686,400 pixels), "typically referred to as 2K", is "considered experimental."

Popular documented sizes include the expected squares and portraits (1024×1024, 1536×1024, 1024×1536), the 2K tier (2048×2048, 2048×1152), and 4K (3840×2160, 2160×3840). The default is auto, which lets the model pick based on the prompt.

Output format is a separate setting from size. The default is png; jpeg and webp are the only other options, and neither the guide nor its spec contains the word "vector" or "SVG" anywhere. A raster model has no route to a scalable format without a separate step outside the API. More detail, including the guide-versus-cookbook discrepancy on the exact edge limit, lives in why an image comes out the wrong size and the GPT Image 2 prompt generator.

What about Stability AI's image sizes?

A fixed pixel count per endpoint, chosen by ratio rather than dimension. Stability's hosted v2beta endpoints (Ultra, Core, SD3) expose only aspect_ratio, a nine-value enum defaulting to 1:1 — no width or height field exists. Ultra's own documentation states its output is 1 megapixel, defaulting to 1024×1024; Core is documented at 1.5 megapixels. Neither steps nor sampler is exposed on any hosted v2beta endpoint; both live only on Stability's older v1 REST API.

Should you generate large, or upscale after?

Generate close to what the model actually publishes as native, then treat upscaling as its own separate step rather than a bigger version of the same request.

The evidence is scattered across every vendor above, and it points one direction. Black Forest Labs ties erase-v1 quality directly to a ~1 megapixel training resolution. Midjourney's own upscalers double an existing SD image rather than generating natively at HD in one pass, and its docs warn: "Some aspect ratios may be slightly changed when upscaling." That caveat is scoped to the upscale step, not generation. Ideogram 4.0's ceiling nearly quadrupled from 3.0, but that came from a change in native generation, not from upscaling a 3.0 output. None of the vendors here document more requested pixels producing a sharper image; the failure mode is quality loss, not an error.

The practical version: pick the native size or ratio your model actually publishes, generate there, and reach for a dedicated upscaler once you have the composition you want. Requesting 8K in a text prompt is a caption, not a setting, on every surface checked here.

Copy-paste settings that get the size right, per tool

Midjourney, web or Discord:

a quiet harbor at dawn, long exposure, muted color --ar 16:9 --hd

Nano Banana Pro / Nano Banana 2 (Interactions API, response_format):

{
  "type": "image",
  "aspect_ratio": "16:9",
  "image_size": "4K"
}

FLUX.2 (flux-2-pro, flux-2-max, or flux-2-flex):

{
  "prompt": "a quiet harbor at dawn, long exposure, muted color",
  "width": 1536,
  "height": 1024
}

FLUX.1 [dev] or FLUX1.1 [pro] (width and height must be multiples of 32):

{
  "prompt": "a quiet harbor at dawn, long exposure, muted color",
  "width": 1024,
  "height": 768
}

Ideogram 3.0 (aspect_ratio and resolution are mutually exclusive):

aspect_ratio=16x9

Ideogram 4.0 (resolution is the only sizing field):

resolution=2048x1152

GPT Image 2:

{
  "model": "gpt-image-2",
  "prompt": "a quiet harbor at dawn, long exposure, muted color",
  "size": "1536x1024"
}

Sources and dates

Fetched and read directly, not summarized, on September 2, 2026 unless marked otherwise:

• Midjourney Image Size & Resolution, Version (updated Sep 1, 2026), Aspect Ratio, and Upscalers — fetched via the Zendesk Help Center JSON API, since docs.midjourney.com 403s to a direct fetch.
• Google Gemini API image generation guide and generateContent API reference (ImageConfig object) — fetched with a cookie jar plus ?hl=en, since a plain request redirects into an OAuth loop.
• Black Forest Labs api.bfl.ai/openapi.json and docs.bfl.ai/llms-full.txt.
• Ideogram developer.ideogram.ai/openapi.json.
• OpenAI image generation guide.
• Stability AI API reference — figures reused from a verified pass on August 27, 2026; every re-fetch attempted on September 2, 2026 returned a Cloudflare challenge page instead of the documentation.
• Prompt Architects FAQ and pricing, checked September 2, 2026, for the free-tier and Image Prompt Generation plan gating referenced above.

Known unknowns, stated rather than guessed: Midjourney publishes pixel dimensions for four aspect ratios only. FLUX.2 states a 4MP ceiling in prose but no maximum in its request schema. Stability's own docs were unreachable at the time of writing, so its figures are one verified pass old, not same-day.

Prompt Architects does not render any of the pixels in this piece. What it does is turn a rough brief into a structured prompt, with the size, ratio and framing decisions made explicit before you send it to whichever tool above you are using. Image Prompt Generation is on the Pro plan and up; the free plan covers 5 prompt enhancements a day, forever.