Midjourney vs Stable Diffusion vs Flux vs Ideogram vs Leonardo AI: AI Image Generation (2026)

There are five mainstream ways to generate images with AI in 2026, and they split cleanly by architecture of ownership rather than by output style. Two are open-weight ecosystems you can run on your own hardware — Stable Diffusion fully, FLUX partially, variant by variant. Three are platforms: Midjourney (closed entirely), Ideogram (a platform that published downloadable weights in June 2026, with a catch), and Leonardo AI (closed, now a Canva subsidiary).

Every one of the five claims excellent image quality and legible text rendering. Those claims are marketing, and they converge. What does not converge is the license fine print: who owns the output, at what revenue threshold ownership changes, and which model files you may legally deploy for money. That fine print is where this comparison spends most of its time, because it is the part that produces real consequences — a $1M-revenue clause in Midjourney's terms, a free-tier clause in Leonardo's terms that assigns your images to Leonardo, and a widely repeated but wrong claim that Ideogram 4.0's weights are Apache-licensed.

All figures below come from the official pricing pages, terms of service, model licenses and update feeds of the five vendors, accessed 21 July 2026. Where a vendor makes a quality claim, it is labelled as the vendor's own claim. More comparisons live in our guides section.

Midjourney: the closed community platform

Flagship. V8.1 has been the default model since 11 June 2026. Midjourney's own release note says it "better adheres to detailed prompts, and renders text better than ever," generates in roughly 4 seconds standard and 12 seconds in HD, and its HD mode outputs at twice the size and four times the resolution of V7 images. The V8 alpha that preceded it (17 March 2026) introduced native 2K generation via --hd and new parameters including --chaos, --exp and --q 4. A V1 video model has existed since June 2025: image-to-video clips of about 5 seconds, extendable roughly 4 seconds at a time up to four times, at about 8x the cost of an image job.

Pricing. Four paid tiers: Basic $10, Standard $30, Pro $60, Mega $120 per month, with 20% off if you pay a year upfront. The metered unit is Fast GPU hours (3.3 h/month on Basic up to 60 h on Mega, top-ups at $4/hour), not credits or image counts. Standard and above get unlimited "Relax" (queued) image generation; Pro and Mega add unlimited relaxed SD video and Stealth mode for private generations. The official plans chart lists no free tier.

Where it runs. Closed SaaS only — the web app and Discord. No public API and no downloadable weights appear anywhere on Midjourney's docs or pricing pages.

Ownership. The terms of service say you own the assets you create "to the fullest extent possible under applicable law," and ownership survives downgrading or cancelling. Two catches. First, the clause this guide flags hardest: a company (or employee of one) with more than $1,000,000 USD a year in revenue must be subscribed to a Pro or Mega plan to own its assets. Second, you grant Midjourney a perpetual, worldwide, royalty-free, irrevocable license to your prompts and generated assets, and generations are public and remixable by default — only Stealth mode, available on Pro and Mega, keeps them unpublished, and even then on a "best efforts" basis.

Company. Midjourney currently describes itself on its homepage as "a community-funded research lab of 60 people" — an independent, self-funded framing that is Midjourney's own current statement.

Stable Diffusion: the most permissive full stack

Flagship. Stability AI's image flagship is still the Stable Diffusion 3.5 family announced 22 October 2024: SD 3.5 Large (8B parameters, 1 MP output), Large Turbo (distilled, 4-step generation) and Medium (2.5B parameters, 0.25–2 MP, built for consumer hardware). The API adds SD 3.5 Flash and Stable Image Ultra, a hosted service built on Large. Notably, Stability's 2026 announcements have leaned partly outside imaging (Stable Audio 3.0 appears on the API pricing page) — the image line has not had a new foundation release since October 2024.

Pricing. There is no consumer subscription app. Stability sells a developer API metered in credits (1 credit = $0.01, 25 free credits to start): SD 3.5 Large about $0.065 per image, Medium $0.035, Stable Image Ultra $0.08, with edit operations at $0.04–0.08 and upscalers from $0.02 to $0.60.

Where it runs. Everywhere. Weights are downloadable from Hugging Face (gated behind license acceptance) for full self-hosting, and the same models are available through Stability's API and third-party hosts.

License. The Stability AI Community License is the cleanest deal of the five for small operators: free for research, non-commercial use, and commercial use if you or your organisation has under USD $1M in annual revenue. Above that you need an Enterprise License. Outputs are unambiguously yours — the license FAQ states you own outputs generated from the core models or your fine-tunes and can use them at your discretion. The one restriction: you may not use the models to build competing foundation models (fine-tunes and LoRAs are fine).

Company. After a turbulent stretch, Stability took new investment in June 2024 from Greycroft, Coatue, Sound Ventures, Lightspeed and others; Prem Akkaraju (formerly CEO of Weta Digital) was appointed CEO and Sean Parker executive chairman at that time.

FLUX: current-generation open weights, with a license ladder

Flagship. Black Forest Labs — founded by original Stable Diffusion researchers, per wide reporting — shipped FLUX.2 on 25 November 2025, then FLUX.2 klein on 15 January 2026. The lineup matters because the license changes per variant: FLUX.2 [pro], [flex] and [max] are closed API models; FLUX.2 [dev] is an open-weight 32B model; klein comes in 4B and 9B sizes with unified generation-plus-editing, sub-second inference and consumer-GPU targets (the 4B models run in roughly the 13 GB VRAM class per the official post, with FP8/NVFP4 quantizations). FLUX.2 outputs up to 4 MP and accepts up to 10 reference images; BFL's own claim covers "complex typography" and legible fine text.

Pricing. No subscriptions — the API is metered per megapixel: FLUX.2 [max] $0.07 for the first MP, [pro] $0.03, [flex] $0.05, klein 9B $0.015 and klein 4B $0.014, which makes klein the cheapest current-generation model API in this comparison (only Stability's legacy SDXL endpoint, at roughly $0.009, prices lower). Separately, BFL sells commercial self-hosting licenses in tiers (Builder, Platform, Professional, Enterprise, plus a Synthetic Data tier for training-data rights); none of the tier prices are published on the pricing page.

License — read this per variant, not per brand. FLUX.1 [schnell] and FLUX.2 [klein] 4B are Apache 2.0: unrestricted commercial use, including self-hosted commercial deployment. FLUX.1 [dev], FLUX.2 [dev] and FLUX.2 [klein] 9B sit under the FLUX Non-Commercial License for the weights — but per both dev model cards, "generated outputs can be used for personal, scientific, and commercial purposes." In other words: the images are commercially usable; running the model itself as a commercial service requires one of the paid self-host tiers. The FLUX.2 [dev] license also requires NSFW and IP-infringement filtering at input and output. The pro/flex/max models are commercial via the paid API.

Company. BFL lists a $300M Series B at a $3.25B valuation dated 1 December 2025 on its official blog; CEO is co-founder Robin Rombach. Its site footer carries ISO 27001:2022 and SOC 2 Type II badges — relevant if you are buying for an enterprise.

Ideogram: text-in-image specialist, now open-weight but not open-license

Flagship. Ideogram 4.0, released 3 June 2026, is a 9.3B-parameter flow-matching diffusion transformer generating natively up to 2048 px, and it is the first Ideogram foundation model you can download (fp8 and nf4 quantizations on Hugging Face).

The correction that matters. A claim circulating widely in secondary coverage says Ideogram 4.0 is Apache 2.0. It is not. The Hugging Face model cards for both quantizations list the license as "Ideogram 4 Non-Commercial." You may download and run the weights for research and personal use; commercial self-hosting requires a separate paid license from Ideogram ("a license that matches your scale," per the official 4.0 page). If you saw the Apache claim somewhere, it referred at most to inference code, not the model.

Capability. Text rendering inside images is the house specialty, and 4.0 extends it into design-tool territory: bounding-box layout control via normalized 0–1000 coordinates in structured JSON prompts, hex color-palette conditioning, and multilingual text. Ideogram's own benchmark for the earlier 3.0 model claimed an ELO of 1132 against Imagen3 at 1023 and Flux Pro at 998 — a self-run evaluation, so treat it as a vendor claim, not an independent result.

Pricing. The web app has a free tier (free credits at sign-in, public generation only), then Plus at $15/month and Pro at $42/month on annual billing, and Team at $20/user/month annually. Monthly-billing prices were not verifiable this session, so we do not quote them — see the pricing page linked in Sources. The API charges per image: 4.0 at $0.03 (Turbo), $0.06 (Default) and $0.10 (Quality).

Ownership. Clean for platform users: "We do not claim ownership of your generated images, and you're free to use them for any purpose, including commercial use," per the official pricing FAQ. Free-tier generations are public. Payments are Stripe card-only — Visa, Mastercard, Amex; no PayPal or wallets. The founders are ex-Google researchers, per wide reporting.

Leonardo AI: the Canva-ecosystem all-rounder

Flagship. Lucid Origin is Leonardo's current in-house model, available on every plan including free, with native Full HD (1920×1080) output and — its own claim — "clean, legible text rendering." The platform now treats video as core: Motion 2.0 plus third-party Hailuo and Wan models, with unlimited relaxed video on the Ultimate tier.

Pricing. Free gets 150 fast tokens/day with public-only creations. Essential is $12/month (and is stated on the tier card as included with Canva Business), Premium $30 adds unlimited relaxed image generation on select models, Ultimate $60 adds unlimited relaxed video; all ex. tax. Team plans start at $24/seat/month with a 3-seat minimum. The API is pay-as-you-go with a $5 starting credit.

Ownership — the sharpest split in this comparison. Leonardo's terms (last updated 19 January 2026) draw a line no other vendor here draws. For paid subscribers, clause 8.3 says IP in the content you create "will vest in you upon creation." For free subscribers, clause 8.7 says IP in outputs "will vest in US" — that is, in Leonardo. Free-tier users assign their images to the company at the moment of creation. On top of that, public content carries a perpetual, transferable license to Leonardo including AI-training and "any commercial purpose"; private content on paid tiers will not be used for training without express written consent. If you use Leonardo for anything you might later sell, the free tier is not a trial of the paid deal — it is a different deal.

Company. Leonardo (an Australian entity, Leonardo Interactive Pty Ltd) was acquired by Canva; Canva's newsroom says the roughly 120-person team continues to operate independently and that Leonardo's Phoenix model feeds Canva's Magic Studio. No open weights, closed SaaS plus API only.

License and ownership, side by side

This is the load-bearing table. Everything in it is from the vendors' own terms, licenses and model cards, accessed 21 July 2026.

Which one should you pick

The evidence in this guide points to different winners for different buyers, not one winner.

• You want a polished closed platform and a community around it — a paid Midjourney tier. Unlimited relaxed generation from $30/month and native 2K HD on the current V8.1 default are the concrete draws. Check the $1M-revenue clause first if you work at a company of any size, and budget for Pro if privacy (Stealth) matters.
• You want open weights, full control and commercial self-hosting — Stable Diffusion 3.5 if you are under $1M revenue and want the most permissive full stack, or FLUX.2 klein 4B if you want a current-generation (January 2026) model under Apache 2.0 with no revenue test at all. FLUX also happens to have the cheapest quality API here ($0.014–0.03 per ~1 MP image) if you would rather not run GPUs.
• Your work is text-heavy — posters, UI mocks, packaging, multilingual layouts — Ideogram's platform or API. Its layout control (bounding boxes, hex palettes) is unique in this group; its quality lead is its own claim. Use the app or API for commercial work; the downloaded weights are non-commercial.
• You live in the Canva ecosystem and want one canvas for image, video and custom models — a paid Leonardo tier, and specifically paid: the free tier assigns output ownership to Leonardo.

The ASEAN angle, honestly

None of the five vendors offers regional pricing — every official pricing page is USD-denominated, and Leonardo's prices are ex. tax, so Singapore GST lands on top. Ideogram takes Stripe card payments only (Visa, Mastercard, Amex — no PayPal or wallets), which matters in markets where cards are not the default. Leonardo is the one APAC-headquartered option, as an Australian company under Canva. The strongest regional argument, though, is data control: for PDPA-sensitive work, self-hosting SD 3.5 or the Apache-licensed FLUX weights (klein 4B, schnell) is the only way to keep prompts and outputs entirely in-country, because every closed platform here processes your inputs on its own infrastructure.

FAQ

Can I use the images commercially?

Usually, with per-product traps. Midjourney: yes on any paid plan, but companies over $1M annual revenue must be on Pro or Mega to own their assets. Stable Diffusion 3.5: yes, outputs are yours; the model itself is free commercially only under $1M revenue. FLUX: outputs are commercially usable from every variant, including the non-commercial-weight dev models. Ideogram: yes for platform-generated images ("any purpose, including commercial"); not from self-hosted weights without a paid license. Leonardo: yes on paid tiers; on the free tier the outputs vest in Leonardo, not you.

Can I run any of these on my own GPU?

Three of the five publish weights. SD 3.5 (Large, Large Turbo, Medium) is fully self-hostable, with Medium aimed at consumer hardware. FLUX.2 dev (32B) and klein (4B/9B) are downloadable — the 4B klein models target roughly the 13 GB VRAM class per BFL's release post, and only the 4B and FLUX.1 schnell are Apache-licensed for unrestricted use. Ideogram 4.0 (9.3B, fp8/nf4) is downloadable for non-commercial use. Midjourney and Leonardo publish no weights.

Which renders text inside images best?

No independent benchmark in our sources settles this, so label everything as vendor claims: Ideogram has made text-in-image its specialty since launch and cites its own ELO evaluation (1132 for Ideogram 3.0 vs 998 for Flux Pro) plus 4.0's bounding-box text layout; Midjourney says V8.1 "renders text better than ever"; BFL claims legible fine text and complex typography for FLUX.2; Leonardo claims clean, legible text for Lucid Origin. Ideogram is the only one that ships layout-level text controls (JSON bounding boxes, hex palettes), which is a verifiable feature rather than a quality claim.

Do free tiers exist?

Ideogram and Leonardo have real free web tiers — Ideogram's is public-generation-only, and Leonardo's comes with the ownership catch above. Stability gives 25 free API credits (a handful of test images), and BFL has a demo playground but no free consumer app. Midjourney's official plans chart lists only paid tiers; we could not verify any current free-trial availability. If a free tier matters commercially, Ideogram's is the only free consumer web tier here whose outputs are both yours and commercially usable — though they are public. (Stability's 25 free API credits also produce outputs you own, but that is a developer trial, not a web app.)

Want the traction data behind these entries? Browse the AI directory sorted by traction — noting that for this mixed open/closed field, star counts only describe the open-source side.