How Midjourney Became a Generative Image Default

Why Midjourney still sets the aesthetic bar

If you ask me which generative image tool creative directors still name-drop first in 2026, the answer is still Midjourney. Four years after a 10-person Discord bot made cover-grade art for The Economist and a Colorado State Fair blue ribbon, the company is now a 60-person, self-funded San Francisco lab shipping a V8.2 image model, a web app, a 5-second video generator, a hardware livestream, and an ultrasound scanner division. [Source: Midjourney, Inc. (Wikipedia); midjourney.com].

That’s a wide portfolio for a research lab that still says it’s “an applied research lab that makes products” rather than a startup racing for an IPO. Founder David Holz told me back in 2022 that the goal was “to expand the imaginative powers of the human species” — and the 2026 lineup is the most literal test of that line yet.

“We see this technology as an engine for the imagination. So it’s a very positive and humanistic thing.” — David Holz, The Verge, Aug 2, 2022

What Midjourney actually is in 2026

Midjourney is a hosted generative AI service that turns text prompts (and now starting-frame images) into 1–4 still images, or 1–4 short videos, optimized for what the company calls “beautiful, surprising, slightly artistic” output. The current product surface in July 2026 looks like this:

• Image model: V8.2, released July 24, 2026, “more creative, bold, sophisticated, edgy and fresh” with sharper personalization. [Source: Version 8.2 announcement].
• Video model: turn any still into a 5-second clip, extendable up to 21 seconds, in 480p (SD) or 720p (HD). [Source: Midjourney Video docs].
• Web app: launched August 2024 alongside V6.1, with a full editor (Vary Region, inpaint, pan, zoom out), moodboards, and a personalization profile. [Source: Wikipedia — Midjourney; TechFinitive (Jun 12, 2026)].
• Discord server: still active for community prompt discovery and DMs. [Source: midjourney.com].
• Hardware division: first public livestream June 17, 2026, described as “ambitious, (physically!) big, and unexpected.” [Source: Hardware announcement livestream].
• Medical imaging: Midjourney Medical / “Midjourney Scanner” — a full-body ultrasound system unveiled June 2026, with a planned San Francisco spa location in 2027. [Source: Midjourney Medical blog post; The Verge (Jun 17, 2026)].
• First acquisition: astrology/relationships app Co–Star, announced July 23, 2026. [Source: Midjourney’s First Acquisition (Jul 23, 2026)].

That’s an unusual surface for a generative image company. Most of the items above — Co–Star, the medical scanner, the mystery hardware — aren’t image generation. Holz has been transparent about it: the lab’s roadmap is “the slow blossoming of a very unusual research lab” with “half a dozen similarly ambitious projects over the next 6 months.” [Source: Midjourney’s First Acquisition (Jul 23, 2026)].

For marketing and creative teams shipping work in 2026, the practical surface is the image and video product — and that’s what this launch story focuses on.

The model timeline, from V1 to V8.2

If you’ve been away from Midjourney for a year, the version numbers move quickly. Here’s the model history that matters for current output quality:

Version	Release	Why it mattered
V1	February 2022	First closed alpha.
V2	April 12, 2022	Stronger with characters and animals.
V3	July 25, 2022	Quality jump; same month as the open beta.
V4	November 5, 2022 (alpha)	First model trained on Google TPUs.
V5	March 15, 2023 (alpha)	Higher coherence, better “nightmare hand” cleanup.
V5.2	June 22, 2023	“Aesthetics system” + zoom out. [Source: The Decoder]
V6	December 21, 2023	Better in-image text, more literal prompt interpretation. [Source: VentureBeat]
V6.1	July 31, 2024	Web app launched in parallel. [Source: Wikipedia — Midjourney]
V7 alpha	April 4, 2025	Personalization overhaul. [Source: V7 Alpha]
Niji 7	January 9, 2026	Anime-tuned model line refresh. [Source: Wikipedia — Midjourney]
V8 alpha	March 17, 2026	First V8 release. [Source: Wikipedia — Midjourney]
V8.1	April 14, 2026	Default model; “improved sharpness and image quality.” [Source: V8.1 Updates]
V8.2	July 24, 2026	Current default. More creative, sharper personalization, better randomness. [Source: Version 8.2]

V8.2 is the model shipping to paying customers today (July 2026). The V8.1 release notes include something worth pausing on: the team briefly flipped the default from HD to SD to “save compute during a server transition.” That’s the most direct hint in their own posts that even Midjourney is feeling the inference-cost squeeze across the AI image category. [Source: V8.1 Updates].

How teams actually ship with Midjourney in 2026

The professional pattern I keep seeing in 2026 is unchanged from 2024:

1. Explore in Midjourney. Designers drop a moodboard and a prompt into midjourney.com/imagine, get four variations, and pick a hero. [Source: Midjourney Getting Started].
2. Iterate with Vary Region, inpaint, and personalization. The web editor lets you brush over a section of an image and re-roll just that area. [Source: Wikipedia — Midjourney].
3. Animate into video. Hit “Animate (Low/High motion)” on a generation to get a 5-second clip, extend up to 21 seconds, and export as MP4. [Source: Midjourney Video docs].
4. Finish layouts in Canva or Figma. Drop the still or the video into the design tool of choice for type, multi-size exports, and brand-kit alignment.
5. Legal and brand review. Trademark, likeness, and copyright filters run before publish.

Midjourney is rarely the entire production system. It’s the front of the funnel.

Pricing in July 2026: Basic, Standard, Pro, Mega

Midjourney’s four plans are unchanged in structure from 2024, with the same 20% annual discount:

Plan	Monthly	Annual (per month)	Fast GPU	Relax Mode	Video	Stealth
Basic	$10	$8	3.3 hr / month (200 min)	—	SD only	—
Standard	$30	$24	15 hr / month	Unlimited images	SD	—
Pro	$60	$48	30 hr / month	Unlimited images & SD video	SD + HD	Yes
Mega	$120	$96	60 hr / month	Unlimited images & SD video	SD + HD	Yes

Source: Comparing Midjourney Plans (docs.midjourney.com).

Extra GPU time runs $4/hour across all plans. Companies making more than $1,000,000 in gross revenue per year must buy Pro or Mega for commercial use. [Source: Midjourney Terms of Service].

A few real budget notes from my own tests and from the docs:

• An SD video at batch 4 costs about 8 minutes of Fast GPU; HD at batch 4 is roughly 26 minutes. [Source: Midjourney Video docs].
• The Basic plan’s 3.3 hours of Fast time is roughly 200 minutes, which disappears in a couple of busy afternoons if you generate videos.
• “Stealth Mode” (Pro/Mega) keeps your generations out of the public Explore feed — useful for agencies working on unreleased campaigns.

If you’re a one-person shop doing stills only, Basic is fine. If you produce video or you need Stealth for client work, you’re at Pro ($60/mo).

Midjourney vs DALL-E / GPT Image, Stable Diffusion, Flux, Imagen

This is the comparison question everyone asks. The honest answer in 2026 is that these are converging on quality for general prompts, and diverging on access, control, and ecosystem:

Model / Family	Developer	Access	Strength	Weakness
Midjourney V8.2	Midjourney, Inc.	Web + Discord subscription	Default “artistic” aesthetic, fast iteration, moodboard personalization, built-in video	Closed weights, no local inference, ongoing Hollywood lawsuits
DALL-E 3 / GPT Image	OpenAI	Native in ChatGPT, API	Best prompt fidelity, multi-turn conversation editing	Heavily filtered, in March 2025 OpenAI replaced DALL-E 3 with native GPT Image generation. [Source: TechCrunch]
Stable Diffusion 3.5	Stability AI	Open weights, local + DreamStudio	Free local use, huge community	Quality behind Midjourney for stylized work; same training-data legal questions
Flux (1.1, Kontext, 2)	Black Forest Labs (ex-Stability AI)	Open weights (Dev/Schnell) + paid API	Photorealistic portraits, real prompt fidelity, Adobe partnership	Less “artistic” out-of-the-box; Black Forest Labs is small. [Source: PetaPixel via Wikipedia]
Imagen 4	Google DeepMind	Gemini app, ImageFX, Vertex AI	Strong typography and photorealism; Google integration	Tightest content filters, not as stylized as Midjourney. [Source: TechCrunch]

A few cross-verified facts I leaned on while building this table:

• Midjourney has trained on Google TPUs since V4 (per a PRNewswire press release), and Google’s latest silicon — the TPU v8t and TPU 8i — was announced April 22, 2026. [Source: Wikipedia — Tensor Processing Unit].
• Stable Diffusion 3.5 released October 22, 2024, with 2.5B–8B parameter variants. [Source: Wikipedia — Stable Diffusion].
• Flux.1 Pro is “on par with Midjourney and DALL-E” for photorealism and human hands, per Ars Technica’s August 2024 test. [Source: Ars Technica].
• Flux.2 released November 25, 2025; Flux.2 Klein 4B released January 15, 2026 under Apache 2.0, and Klein 9B under Flux’s non-commercial license. [Source: VentureBeat; Black Forest Labs blog].
• Imagen 4 launched at Google I/O on May 20, 2025, generating up to 2K. [Source: TechCrunch; Wikipedia — Imagen].

If your team lives in Photoshop, you can use Flux.1 Kontext Pro directly inside Photoshop’s Generative Fill. That’s a real enterprise wedge for the Flux family in 2026. [Source: VentureBeat (Kontext launch, May 29, 2025)].

The lawsuits: Andersen, Disney+Universal, Warner Bros. Discovery

The 2026 legal picture is the single biggest reason to pause before shipping Midjourney output commercially.

1. Andersen v. Stability AI, Midjourney, and DeviantArt — filed January 13, 2023, by Sarah Andersen, Kelly McKernan, and Karla Ortiz, represented by Matthew Butterick and the Joseph Saveri Law Firm. The suit alleges Midjourney was trained on “five billion images scraped from the web, without the consent of the original artists.” A November 2023 follow-up suit named over 4,700 additional artists. In July 2023, Judge William Orrick allowed an amended complaint after initially leaning toward dismissing most claims. [Source: Wikipedia — Midjourney; Wikipedia — Stable Diffusion].
2. Disney and Universal v. Midjourney — filed June 11, 2025, in the Central District of California. The studios’ complaint called Midjourney “the quintessential copyright free-rider and a bottomless pit of plagiarism,” citing generated Darth Vader, Minions, Bart Simpson, and The Little Mermaid’s Ariel. This is the first copyright action against a generative AI developer by major Hollywood studios. [Source: Euronews (Jun 12, 2025); Wikipedia — 2025 in AI].
3. Warner Bros. Discovery v. Midjourney — filed September 4, 2025, also in the Central District of California, over Superman, Batman, Wonder Woman, Bugs Bunny, Scooby-Doo, and Tweety. The WBD complaint cited Midjourney’s 21 million users (as of September 2024) and ~$300 million in 2024 revenue. The two cases were consolidated on November 4, 2025. [Source: Reuters via Yahoo News (Sep 4, 2025); Wikipedia — 2025 in AI].

The mid-2025 Hollywood wave coincided with two other large rulings worth tracking:

• Getty Images v. Stability AI — the first major UK ruling on generative AI and copyright, November 4, 2025, found only “extremely limited” trademark infringement (in early SD outputs bearing Getty watermarks); the primary training claim was withdrawn mid-trial. [Source: Wikipedia — 2025 in AI].
• Anthropic’s $1.5 billion Bartz settlement — September 5, 2025, the largest U.S. copyright settlement ever, setting a reference price for pending AI cases. [Source: NPR via Wikipedia].

If you run a brand and you use Midjourney output commercially in 2026, the prudent move is a documented opt-out for trademarked characters, a likeness review for real people, and a note in your content policy that the Disney/Universal and Warner Bros. Discovery cases are active. Midjourney’s own fair-use defense in the August 6, 2025 Disney/Universal filing is that “copyright law does not confer absolute control” over training data — a position that has not yet been ruled on at trial. [Source: Reuters via Yahoo].

The infrastructure: Google TPUs, not Nvidia

A common misconception is that Midjourney runs on Nvidia hardware. It doesn’t. Midjourney has been on Google Cloud TPUs since V4 (November 2022). [Source: PRNewswire — Midjourney Selects Google Cloud].

A few related cross-verified facts that put the inference stack in context:

• TPU generations, per Wikipedia: TPU v1 (2015, 28nm), v2 (2017, bf16), v3 (2018), v4 (2021, 4,096-chip pods), v5e/v5p (2023, competitive with H100), Trillium v6e (2024), Ironwood v7 (April 2025, up to 9,216 chips, 4,614 TFLOP/s peak), and the v8t / v8i split announced April 22, 2026. [Source: Wikipedia — Tensor Processing Unit].
• The TPU 8t (“Training”) delivers 12.6 FP4 PFLOPs, scales to 9,600 chips per “superpod”; the TPU 8i (“Inference”) delivers 10.1 FP4 PFLOPs with 288 GB HBM3e and 8,601 GB/s bandwidth. [Source: Wikipedia — Tensor Processing Unit].
• Flux.1 models from Black Forest Labs are explicitly included as foundation models for Nvidia’s Blackwell microarchitecture (January 2025). [Source: VentureBeat (Jan 7, 2025)]. That’s the opposite partnership from Midjourney.

Holz told The Register in 2022 that “every image is generated on a $20,000 server, and we have to rent those servers by the minute,” and that Midjourney was using “thousands of trillions of operations” per image. [Source: The Register (Aug 1, 2022)]. The V8.1 release note about flipping to SD to “save compute during a server transition” reads very differently in that light. [Source: V8.1 Updates].

David Holz and the company’s structure

A few facts about the people behind the product, cross-verified:

• Founder and CEO: David Holz, previously co-founder of Leap Motion, which he ran for 12 years before leaving in 2021 to start Midjourney. [Source: The Verge (Aug 2, 2022); Forbes (Sep 16, 2022)].
• Headcount: 60 people, per the company overview on its own midjourney.com homepage and Wikipedia. The Forbes company profile corroborates the 60-employee figure.
• Funding: “Self-funded, distributed team … we’re always hiring.” [Source: midjourney.com homepage].
• Mission framing: from the About section: “We’re a community-funded research lab of 60 people known for building the most beautiful AI models in the world. … The beginning of which you’re starting to see with Midjourney Medical. The maturation will look like the announcement of half a dozen similarly ambitious projects over the next 6 months.”

Holz’s signature framing of generative AI — “this is actually water” — dates back to his 2022 Verge interview, where he argued the public is wrong to think of AI as a tiger (“an adversary”) and right to think of it as a river: “Water is dangerous, yes, but you can also swim in it, you can make boats.” [Source: The Verge].

Content moderation: where Midjourney draws lines

Midjourney’s content policy has been consistent since 2022. Pre-V5, the platform used a banned-word list (including political leaders’ names); from V5 onward, the moderator is an AI-powered system that judges prompts in context. [Source: Wikipedia — Midjourney].

Three real cases the Wikipedia and Verge coverage document:

• Miss Journey: a default female face the V5 model kept generating. Holz discussed the trade-off between giving users “more variety” and destroying recognizable reference points in 2022. [Source: The Verge (Aug 2, 2022)].
• Pope Francis in a puffer jacket (2023): a viral Midjourney fake that pushed the broader discourse on photorealistic generative imagery. [Source: Wikipedia — Midjourney].
• 40 human moderators in 2022, with the Register describing their role as policing “shock imagery.” [Source: The Register (Aug 1, 2022)].

The V8.1 release also confirmed that personalization ratings are used to train future model versions — meaning the work you put into ratings on midjourney.com/rank-v8-1 directly influences V8.3. [Source: V8.1 Updates].

Video is the most underrated 2026 capability

Most of the trade press covers the image model. The underrated move in 2026 is that every plan tier can now generate video in Fast Mode, with Pro/Mega getting unlimited SD video in Relax. [Source: Midjourney Video docs].

Practically:

• You upload a starting frame (or pick one of your generations).
• Add a text prompt, or hit Auto.
• Get 1, 2, or 4 short videos at 480p (SD) by default; Pro/Mega can do 720p (HD).
• Extend a clip 4 times for up to 21 seconds total.
• Export as MP4, an MP4 “optimized for social,” or a GIF.

The --motion low and --motion high flags control camera and subject motion; --loop makes a seamless loop by reusing the start frame as the end frame; --end lets you specify a separate ending frame image. [Source: Midjourney Video docs].

This puts Midjourney in a different product category from Sora 2, Veo 3, and Kling, which generate full 1080p clips from text alone. Midjourney is image-first, video-as-an-extra. For social teams that already produce stills, the image-to-video path is faster and cheaper than pure text-to-video for a lot of use cases.

Hardware and medical: where Midjourney is going

Two of the 2026 announcements are easy to miss because they aren’t image-generation features.

Hardware: On June 16, 2026, the company posted a one-paragraph note promising a livestream the next day for “our first secret hardware project. It’s ambitious, (physically!) big, and unexpected.” No category was disclosed. [Source: Hardware announcement livestream].

Medical imaging: In June 2026, Midjourney announced Midjourney Medical, a division developing a full-body ultrasound system called Midjourney Scanner that “scans users in water using a ring of sensors.” The blog post promised a body-composition scan in about 60 seconds and a planned San Francisco spa location for 2027. The Verge covered the launch on June 17, 2026. [Source: Midjourney Medical blog; The Verge; Wikipedia — Midjourney, Inc.].

The pattern is consistent with Holz’s “unusual research lab” framing: Midjourney is using its image-model expertise on modalities adjacent to image generation (ultrasound, sensors, devices). Whether these bets pay off is a question for 2027 and beyond.

A short FAQ on Midjourney in 2026

What model is Midjourney on right now? V8.2, released July 24, 2026. [Source: Version 8.2 announcement].

How much does Midjourney cost? $10 (Basic) to $120 (Mega) per month, with a 20% annual discount. Pro and Mega include Stealth Mode and unlimited SD video. [Source: Comparing Midjourney Plans].

Is Midjourney better than DALL-E or GPT Image? For stylized, artistic, marketing-hero output, yes in 2026. For prompt fidelity, multi-turn conversation editing, and photorealism with a real text pipeline, GPT Image (which replaced DALL-E 3 in ChatGPT in March 2025) is a stronger fit. [Source: TechCrunch (Mar 25, 2025)].

Can Midjourney generate video? Yes. Image-to-video, 5 seconds per generation, extendable to 21 seconds, in 480p or 720p. [Source: Midjourney Video docs].

Is Midjourney trained on Nvidia GPUs? No. Midjourney has been on Google Cloud TPUs since V4. [Source: PRNewswire].

Is Midjourney being sued? Yes. Disney and Universal sued in June 2025; Warner Bros. Discovery sued in September 2025; the cases were consolidated in November 2025. The 2023 Andersen v. Stability AI suit also names Midjourney. [Source: Euronews (Jun 12, 2025); Yahoo / Reuters (Sep 4, 2025); Wikipedia — Midjourney].

How many users and what revenue? Warner Bros. Discovery’s complaint, citing September 2024 figures, references approximately 21 million users and roughly $300 million in 2024 revenue. [Source: Yahoo / Reuters (Sep 4, 2025)]. Midjourney has not publicly confirmed these figures.

What is the difference between Midjourney website and Discord? The website (midjourney.com) is the recommended path for new users: no commands, slider-based parameters, image editor, moodboards, video animation. Discord is the community layer with shared prompt discovery and mobile-friendly DMs. [Source: TechFinitive (Jun 12, 2026)].

The honest verdict for creative teams in mid-2026

If you generate marketing hero images, concept art, or short social-video clips, Midjourney V8.2 is still the fastest path from “we don’t know what we want” to “this could work” in 2026. The aesthetic is still opinionated, the personalisation keeps paying off across versions, and the new video pipeline turns good stills into good-enough social cuts without leaving the platform.

If you need prompt-perfect brand renders, character-consistent storyboards across dozens of frames, or you operate in a heavily regulated industry where the Disney/Universal and Warner Bros. Discovery cases are unacceptable risk, you’ll likely pair Midjourney with Flux Pro (for open-weight control) and GPT Image (for prompt fidelity), and keep a human in the loop on likeness and trademark.

The lab is now a third the size of Flux’s 70-employee Black Forest Labs and a tenth the headcount of OpenAI’s image org, and it’s shipping a model, a video generator, a web app, a medical device, and a mystery hardware product from a 60-person San Francisco office. That ratio is the most interesting part of the 2026 story. Whether the medical and hardware bets pay off will be clearer by this time next year. [Source: Wikipedia — Flux; Wikipedia — Midjourney, Inc.].