Midjourney
Last reviewed
Jun 2, 2026
Sources
No citations yet
Review status
Needs citations
Revision
v5 · 7,896 words
Improve this article
Add missing citations, update stale details, or suggest a clearer explanation.
Suggest edit
CtrlK
Last reviewed
Jun 2, 2026
Sources
No citations yet
Review status
Needs citations
Revision
v5 · 7,896 words
Add missing citations, update stale details, or suggest a clearer explanation.
Midjourney is an artificial intelligence image generation program and independent research lab headquartered in San Francisco, California. Founded in August 2021 by David Holz, the platform uses diffusion models to generate images from natural language text descriptions known as prompts. Since its open beta launch on July 12, 2022, Midjourney has grown into one of the most widely used generative AI tools in the world, reaching more than 21 million registered users on Discord by mid-2025 and generating an estimated $500 million in annual revenue that year.[1][2][3] The company operates without venture capital funding, has been profitable since its second month of operation, and is known for a distinctive cinematic, painterly aesthetic that has made it the preferred tool for many concept artists, designers, and hobbyists.[4][5]
Midjourney, Inc. describes itself as "an independent research lab exploring new mediums of thought and expanding the imaginative powers of the human species." The company was founded by David Holz, who previously co-founded Leap Motion in 2010, a sensor technology company that developed optical hand-tracking devices for virtual and augmented reality. Leap Motion was acquired by British firm Ultrahaptics (now Ultraleap) in May 2019 for roughly $30 million, a fraction of its peak 2013 valuation of $306 million.[6][7] Holz, who holds degrees in applied mathematics and physics from the University of North Carolina and the University of Florida, departed in 2021 after a twelve-year tenure, citing burnout from the demands of a venture-backed environment.[6][8]
After leaving Leap Motion, Holz turned his attention to diffusion models. He has cited Cornell University's research on Denoising Diffusion Probabilistic Models and the release of OpenAI's CLIP (Contrastive Language-Image Pre-training) as the inspiration for the project, both of which signaled that high-quality text-to-image generation was suddenly within reach.[6][8] Holz assembled an initial team of ten engineers in San Francisco and had a working private demo running by September 2021.[6]
Midjourney has maintained a remarkably lean structure throughout its rapid growth. When the open beta launched in 2022, the team consisted of roughly eleven people. By mid-2025 the workforce had grown to between 107 and 163 employees, depending on the source, still small relative to revenue.[9][10] Revenue per employee exceeds $5 million annually, making Midjourney one of the most efficient artificial intelligence companies of the era.[9][11]
Midjourney has never raised external venture capital. Holz has repeatedly declined investor offers, choosing instead to bootstrap the company entirely through subscription revenue. The company reached profitability in August 2022, one month after launching its open beta.[4][11] Revenue grew from an estimated $50 million in 2022 to $200 million in 2023, approximately $300 million in 2024, and roughly $500 million in 2025, all with effectively zero traditional marketing spend and through organic community growth and word of mouth.[5][11][12] In an interview published in March 2026, Holz disclosed that the company was beginning to explore a small hardware project, informally called "Orb," which he conceded might require Midjourney's first outside capital if the program expands.[13]
In 2025, TIME magazine named David Holz to its list of the 100 Most Influential People in AI, citing Midjourney's role in popularizing AI image generation and the company's culture of community-first product design.[2]
David Holz began working on Midjourney in mid-2021, motivated by the rapid advances in text-to-image generation that followed the release of CLIP. CLIP showed that AI systems could effectively assess the alignment between text descriptions and generated images, sparking a wave of research in the field, including OpenAI's DALL-E and the Latent Diffusion Models paper that became the foundation for Stable Diffusion.[6][14] Holz set up an independent lab in San Francisco with an initial team of ten engineers, and the team had a working private demo by September 2021.[6][15]
Midjourney launched a public Discord server in February 2022, and shipped its first model version (V1) the same month. The model produced abstract and painterly images with limited coherence, but it demonstrated the basic capability and built an early community of testers. An invitation-only private beta opened in March 2022. On July 12, 2022, Midjourney announced its open beta, making the tool available to the general public through Discord, with images visible to all users in shared channels.[15][16][17] The platform attracted a one-million-user Discord server within three months of public launch.[4][16]
The open beta attracted users at an unusual pace. By the end of 2022, Midjourney had released its fourth model version and was iterating rapidly on image quality. The platform's social nature, in which every image was visible in public Discord channels by default, created a built-in viral loop: users could see what others were creating, learn from prompts, and share striking results on Twitter and Reddit.[4][17]
In September 2022, the platform received mainstream attention when Jason Allen won first place in the digital arts category at the Colorado State Fair with an image titled "Théâtre D'opéra Spatial," which he had produced using Midjourney. The result ignited widespread debate about AI-generated art and its place in creative competitions.[18][19] In March 2023, a Midjourney-generated image of Pope Francis in a white Balenciaga-style puffer jacket went viral on social media, fooling millions of viewers who believed it was a real photograph. The incident, described as "the first real mass-level AI misinformation case," prompted Midjourney to end its free trial program shortly afterward.[20][21]
Throughout 2023 and 2024 Midjourney continued to improve its models and expand beyond Discord. An alpha web interface at midjourney.com opened to high-volume users (the "5,000 Club") in late 2023, and the full web app launched alongside model version 6.1 in August 2024, allowing users to generate and edit images without a Discord account.[22][23] On April 3, 2025, the company released V7 in alpha; on June 17, 2025, V7 became the default model.[24][25]
On June 18, 2025, Midjourney released V1 Video, its first video generation model, accessed through an image-to-video workflow.[26][27] On July 25, 2025, the company rolled out looping clips, custom start and end frames, and Discord access to video generation.[28] On July 30, 2025, Midjourney launched midjourney.tv, a 24/7 livestream of trending AI-generated videos made by users.[29][30] In August 2025, Meta announced a partnership to license Midjourney's image and video technology for use across Facebook, Instagram, and WhatsApp; Holz publicly reaffirmed that Midjourney would remain independent and investor-free.[31][32]
On March 17, 2026, Midjourney released the V8 alpha at alpha.midjourney.com, with native 2K HD rendering, dramatically faster speeds, and better prompt adherence.[33] On April 30, 2026, V8.1 rolled out to midjourney.com and Discord, becoming the fastest model the company had shipped to that point.[34]
Midjourney uses closed-source, proprietary software based on latent diffusion model architecture. While the company does not publicly disclose its exact model architecture or training data, the general approach is well understood. The system combines a large language model for text encoding with a diffusion model for image synthesis.[35]
The image generation process involves several key stages:
Midjourney's models are trained on a large dataset of images sourced from the internet, including images that overlap with LAION's open-source datasets, particularly LAION-5B.[37][38] The company has also incorporated user feedback data into its training pipeline, using information about which images users upscale, favorite, or select as a reinforcement learning signal to improve aesthetic quality.[35] This human-in-the-loop signal is one explanation for the platform's pronounced house style.
Major model versions starting with V4 have been built on what Midjourney calls its "AI superclusters," and V4 in particular was an entirely new codebase trained on a new Google TPU cluster.[39][40] V6 was the third Midjourney model trained from scratch and took roughly nine months to develop.[39] Holz has framed Midjourney's longer-term goal in terms of real-time, open-world simulation rather than static images, with V1 Video described publicly as a "stepping stone" toward that target.[26][41]
Midjourney provides two primary interfaces for generating images:
| Interface | Description | Availability |
|---|---|---|
| Discord bot | The original interface, accessible through Midjourney's official Discord server or any Discord server where the bot has been added. Users type commands in chat channels to generate images. | Available since February 2022 |
| Web application | A standalone web interface at midjourney.com with a visual prompt builder, image gallery, settings panel, conversation mode, and the integrated image Editor. | Alpha rollout in late 2023 to high-volume users; general availability with V6.1 in August 2024 |
On Discord, users interact with the Midjourney Bot by typing slash commands in text channels. The bot processes the request and returns results directly in the channel. On the web interface, users type prompts into an "Imagine" bar and can adjust settings such as aspect ratio, model version, stylization, and speed through a sidebar panel.[22]
Midjourney offers a range of commands and features for creating and modifying images. The core workflow involves generating an initial set of images, then refining the results through upscaling, variations, region edits, and outpainting.
| Command | Description |
|---|---|
/imagine | The primary command for generating images. Users provide a text prompt, and Midjourney returns a grid of four image previews. |
/blend | Uploads and combines two to five images, merging their concepts, moods, and visual elements into a single blended output. |
/describe | Accepts an uploaded image and returns four text prompts that could reproduce a similar image, useful for reverse-engineering visual styles. |
/shorten | Analyzes a long prompt and identifies the most impactful tokens, suggesting shorter versions that preserve the core meaning. |
/settings | Opens a panel to configure default parameters such as model version, stylization level, and generation mode. |
/subscribe | Directs the user to the subscription management page. |
After generating an initial grid of four images, users can refine their results using a set of modification tools:[42]
| Tool | Description |
|---|---|
| Upscale (Subtle) | Increases the resolution of a selected image while preserving its original details and style closely. |
| Upscale (Creative) | Increases resolution while adding new details, potentially introducing fresh visual elements. |
| Variations (Subtle) | Generates new versions of a selected image with small, detailed changes that stay close to the original composition. |
| Variations (Strong) | Generates new versions with more significant changes while maintaining the overall theme. |
| Zoom Out | Expands the canvas outward in all directions, generating new content around the original image at 1.5x or 2x zoom (outpainting). |
| Pan | Extends the image in a specific direction (up, down, left, or right) while keeping the original content in place. |
| Vary (Region) | Allows users to select a specific region of the image and regenerate only that area, functioning as an inpainting tool. |
The web Editor at midjourney.com provides a single interface for Remix, Vary (Region), Pan, and Zoom Out, with freehand and rectangular selection tools that target the affected area.[42] In V7 and later, the Editor also supports conversational interaction, allowing users to describe changes in natural language (for example, "replace the cat with an owl" or "change the sky to sunset") inside Draft Mode.[24][43]
In addition to text prompts, Midjourney supports image-based reference parameters that lock in specific characters, styles, or moods across multiple generations:
| Parameter | Introduced | Purpose |
|---|---|---|
Style Reference (--sref) | 2024, V6 | Applies the visual character of a reference image (colors, composition, texture, atmosphere) to new prompts. Six versions of the algorithm exist by V7, controlled with --sv. |
Character Reference (--cref) | 2024, V6 | Holds a character's identity (face, costume, key features) consistent across images. Replaced by Omni Reference in V7. |
Omni Reference (--oref) | 2025, V7 | A unified reference system that handles characters, objects, and visual elements from a single image. Weight is controlled with --ow (default 100, range 1 to 1000). |
| Moodboards | 2024 to 2025 | Curated collections of images that train a persistent aesthetic vector for a user or project, combinable with personalization. |
Midjourney supports a variety of parameters that users append to prompts to control the output. Effective use of these parameters is a key part of prompt engineering for visual generation.
| Parameter | Syntax | Range | Description |
|---|---|---|---|
| Aspect ratio | --ar | e.g., 16:9, 3:2 | Sets the width-to-height ratio of the generated image. Default is 1:1. |
| Stylize | --s | 0 to 1000 | Controls how much creative freedom Midjourney has over the prompt. Lower values track the prompt literally; higher values favor aesthetic interpretation. Default is 100. |
| Chaos | --c | 0 to 100 | Determines how varied the four images in each grid are. Higher values push outputs in more divergent directions. |
| Weird | --w | 0 to 3000 | Introduces unusual or quirky aesthetics. Higher values produce increasingly experimental results. |
| Quality | --q | 0.25, 0.5, 1, 2, 4 | Controls compute spent on generation. Higher values increase detail at the cost of more GPU time. --q 4 was introduced with V8 alpha for extra coherence. |
| Version | --v | e.g., 5.2, 6, 7, 8 | Specifies which model version to use. |
| Niji | --niji | e.g., 5, 6, 7 | Switches to the Niji model line, optimized for anime and illustrative styles. |
| Style | --style | raw, cute, expressive, scenic | Adjusts the rendering style. raw reduces Midjourney's default aesthetic bias. |
| No | --no | text | Negative prompting: specifies elements that should be excluded from the image. |
| Seed | --seed | 0 to 4294967295 | Sets a specific seed number for reproducibility. The same seed and prompt produce similar results. |
| Tile | --tile | flag | Generates images that can tile seamlessly. |
| HD | --hd | flag | Generates native 2K resolution images (introduced in V8 alpha; default in V8.1). |
| Exp | --exp | flag | Experimental V8 stylization toggle. |
Effective Midjourney prompts tend to follow several general principles, as advised by the company and its documentation team:[49][50]
--no red.Midjourney has released multiple model versions since its initial launch, each representing significant improvements in image quality, prompt understanding, coherence, and speed. The following table summarizes the main image model versions.[17][25][33][39][52][53]
| Version | Release date | Key improvements |
|---|---|---|
| V1 | February 2022 | First model release. Produced abstract, painterly images with limited coherence. Demonstrated basic text-to-image capability. |
| V2 | April 2022 | Improved coherence and visual clarity over V1, though outputs still had a distinctly abstract quality. |
| V3 | July 2022 | Introduced the --stylize and --quality parameters. Better handling of complex scenes and improved overall fidelity. |
| V4 | November 5, 2022 | Major quality leap. Entirely new codebase and architecture trained on a new AI supercluster. Initial render resolution doubled to 512x512. Brought Midjourney into mainstream attention. |
| V5 | March 15, 2023 | Pushed toward photographic realism with native 1024x1024 output and broader stylistic range. More accurate prompt interpretation, improved hand rendering, and unlimited aspect ratios. |
| V5.1 | May 4, 2023 | Refinement of V5 with stronger default aesthetics and improved coherence. Introduced the --style raw parameter for less opinionated outputs. |
| V5.2 | June 22, 2023 | Further aesthetic improvements, sharper detail, and introduction of Zoom Out (outpainting), /shorten, and High Variation mode. |
| V6 | December 20 to 21, 2023 | Third model trained from scratch on a Midjourney supercluster (nine months of training). Enhanced accuracy for long and complex prompts, doubled token limit to 150, improved coherence, and the ability to render legible text within images. Advanced image prompting and remixing. |
| V6.1 | July 30, 2024 | Refinement of V6 with more coherent images, more precise details and textures, new 2x upscalers, a new personalization model, the --q 2 mode, and approximately 25% faster generation. |
| V7 | April 3, 2025 (alpha); June 17, 2025 (default) | Complete rebuild from scratch. Introduced Draft Mode (10x faster, half cost), default-on personalization, Omni Reference for characters and objects, and conversational editing on web. |
| V8 alpha | March 17, 2026 | Released on alpha.midjourney.com. Native 2K HD rendering via --hd, a new --q 4 quality mode, roughly 5x faster generation, improved prompt adherence (especially for text in quotes), grid view, sidebar settings, and improved conversation mode. |
| V8.1 | April 30, 2026 | Brought V8 features to midjourney.com and Discord. HD by default, 3x faster and 3x cheaper HD mode, image weights returned, new Prompt Shortener, and Moodboard and --sref stability. Requires unlocking the Global V7/V8 personalization profile. |
Alongside its main image line, Midjourney offers the Niji series, developed in collaboration with Spellbrush. The Niji models specialize in anime, manga, and illustrative styles influenced by East Asian aesthetics.[54][55]
| Niji version | Release date | Key features |
|---|---|---|
| Niji 4 | November 2022 | First Niji model, designed for anime-style outputs. |
| Niji 5 | April 2023 | Added --style cute, --style expressive, --style scenic, and --style original sub-styles. |
| Niji 6 | June 7, 2024 | Improved Japanese and Chinese text rendering, finer anime detail (especially eye structure), and improved consistency. |
| Niji 7 | January 9, 2026 | Major boost in coherency. Cleaner, flatter rendering closer to actual anime production style. Improved prompt adherence for specific character designs, eye highlights, and pose specification. Compatible with V7 --sref style codes. |
Midjourney operates on a subscription-based model. As of 2025 the platform no longer offers any free access, having discontinued its free trial in April 2023 following the viral Pope Francis image.[20][56] All plans include commercial usage rights and access to the member gallery.
| Plan | Monthly price | Annual price (per month) | Fast GPU hours | Relax mode | Stealth mode | Max concurrent Fast jobs |
|---|---|---|---|---|---|---|
| Basic | $10 | ~$8 | ~3.3 hours | No | No | 3 |
| Standard | $30 | ~$24 | ~15 hours | Unlimited | No | 3 |
| Pro | $60 | ~$48 | ~30 hours | Unlimited | Yes | 12 |
| Mega | $120 | ~$96 | ~60 hours | Unlimited | Yes | 12 |
Fast GPU time is used for priority processing and provides the fastest generation speeds. Once Fast hours are depleted, users on the Standard plan and above can switch to Relax Mode, which queues jobs at lower priority with variable wait times. Unused Fast hours do not roll over between billing cycles.[56]
Stealth Mode, available only on Pro and Mega plans, prevents generated images from appearing in public galleries on midjourney.com. The feature is important for commercial users who need to keep work confidential.[56] Turbo Mode is available for all subscribers, generating images at approximately double the GPU time cost. Video generation uses approximately 8 times the GPU time of an image job, with each job producing four 5-second clips.[26][27]
On June 18, 2025, Midjourney released its first video model, designated V1 Video. The feature uses an image-to-video workflow: users first generate or upload a still image, then press an "Animate" button to convert it into a short video clip.[26][27]
Key details of V1 Video include:
On July 25, 2025, Midjourney added looping clips (low motion or high motion), custom start and end frames, and an option to extend existing videos by appending new end frames.[28] On July 30, 2025, the company launched midjourney.tv, a 24/7 livestream that displays trending V1 Video clips with their prompts and creator names visible on hover. The channel is accessible at midjourney.tv and on YouTube.[29][30]
Midjourney operates in an increasingly competitive AI image generation market. As of 2025, Midjourney is estimated to hold around a quarter of the global AI image generator market share.[3][57] Each major platform has distinct strengths:
| Platform | Developer | Approach | Key strengths |
|---|---|---|---|
| Midjourney | Midjourney, Inc. | Closed-source, subscription | Aesthetic quality, cinematic and painterly output, strong community, and integrated Omni Reference and Moodboards. |
| DALL-E 3 | OpenAI | Closed-source, integrated with ChatGPT | Precise prompt following, beginner-friendly interface, and tight ChatGPT integration. |
| gpt-image-1 | OpenAI | API and ChatGPT image model | Multi-turn editing, strong text rendering, native conversational generation. |
| Stable Diffusion 3 | Stability AI | Open-weight, locally runnable | Full customization, fine-tuning via LoRA, local processing. Free to run on consumer GPUs. |
| FLUX.1 and FLUX.2 | Black Forest Labs | Open-weight and commercial tiers | State-of-the-art photorealism, strong lighting, used in many third-party products. |
| Adobe Firefly | Adobe | Integrated into Creative Cloud | Commercially safe training data, deep Photoshop and Illustrator integration. |
| Imagen 3 and Imagen 4 | Google DeepMind | Closed-source, available via Vertex AI and Gemini | Strong photorealism and text rendering. |
| Ideogram | Ideogram AI | Subscription | Specializes in legible in-image text and typography. |
Midjourney's competitive advantage lies primarily in aesthetic quality. The platform produces images with a distinctive cinematic, richly detailed look that many users and professionals consider superior for artistic and conceptual work. For tasks requiring precise text rendering, exact spatial control, or photorealistic product photography, some competitors may perform better, although the V8 alpha narrowed those gaps substantially.[33][58]
The Meta partnership announced in August 2025, in which Meta licensed Midjourney's "aesthetic technology" for integration into Facebook, Instagram, and WhatsApp, is widely understood as Meta's attempt to close the quality gap between its in-house image tools and rivals such as Sora, FLUX, and Veo.[31][32]
Midjourney has had an outsized impact on the creative landscape since its launch. The platform has been adopted by a wide range of users, from amateur hobbyists exploring their imagination to professional architects, graphic designers, concept artists, illustrators, and advertising agencies using it to prototype visual ideas.[4][59]
Architects and interior designers use Midjourney to visualize design concepts before committing to detailed plans. Graphic designers use it to generate initial mood boards and explore creative directions. Game studios and film production companies have used it for concept art during pre-production. Advertising agencies have adopted the tool for rapid ideation, generating dozens of visual concepts in the time it would traditionally take to produce a single sketch.[4][59]
Midjourney's Discord-based community has fostered a culture of collaborative experimentation. Public channels function as a live gallery where users observe each other's prompts and results, learn techniques, and share discoveries. This social dimension has been a primary factor in the platform's growth and in the broader popularization of AI-generated art.[17] Holz has framed Midjourney's design around community throughout, telling TIME in 2025 that user creations being public by default is core to the product.[2]
The platform has been used in music videos and television production. Hungarian DJ and animator Shane 54 created an entire music video from Midjourney-generated frames. Several magazines, including The Atlantic and The Economist, have used AI-generated artwork on their covers. Midjourney TV, launched in July 2025, surfaces and continuously plays user-generated short video clips, giving a window into the volume and diversity of work being produced on the platform.[29][30]
The rise of Midjourney and similar tools has sparked significant debate within the art community. Some artists have embraced the technology as a powerful new creative tool. Others have voiced deep concern about its potential to reduce demand for human illustrators and designers. Cartoonist Matt Bors captured this tension in 2022, stating: "To developers and technically minded people, it's this cool thing, but to illustrators it's very upsetting because it feels like you've eliminated the need to hire the illustrator."[59] In January 2024, an alleged "Midjourney Style List" containing the names of more than 16,000 artists circulated online and later surfaced in court filings, reigniting the debate over consent and training data.[60]
Holz has consistently framed Midjourney as a tool for human imagination rather than a replacement for human creativity. In a 2022 interview he said, "We like to say we're trying to expand the imaginative powers of the human species. The goal is to make humans more imaginative, not make imaginative machines, which I think is an important distinction." In the same era he described Midjourney as "an engine for the imagination," and he has framed Midjourney's longer-term goal in terms of real-time open-world simulation: "In 10 years, you'll be able to buy an Xbox with a giant AI processor, and all the games are dreams."[8][61]
In August 2022, Jason Allen entered an image titled "Théâtre D'opéra Spatial" in the digital arts category at the Colorado State Fair's fine art competition. The image, which depicted classical figures in an ornate hall looking through a circular viewport, won first place in the "emerging artist" (non-professional) division on August 29 and a $300 prize. Allen had used Midjourney to generate the image, employing at least 624 text prompt iterations before editing the result in Adobe Photoshop and upscaling it with Gigapixel AI.[18][19]
The award drew widespread controversy when it became public that the winning entry was AI-generated. Many traditional artists accused Allen of cheating and argued that AI-generated art should not compete alongside human-created work. The two judges later said they did not know Midjourney used AI but would have awarded Allen the top prize regardless. Allen maintained that he was using Midjourney as a creative tool, similar to how other artists use digital software.[18][19] Starting in 2023, the Colorado State Fair required participants to disclose whether they used AI.[18]
The controversy deepened when Allen applied for copyright protection for the image. On September 5, 2023, the U.S. Copyright Office Review Board issued a final determination that "Théâtre D'opéra Spatial" could not receive copyright registration, concluding that Allen's contribution (typing text prompts) was insufficient to establish human authorship.[62] In 2024, Allen filed an appeal in U.S. District Court in Colorado, challenging the Copyright Office decision.[62]
In March 2023, a Midjourney-generated image of Pope Francis wearing a large white Balenciaga-style puffer jacket went viral across social media platforms, fooling millions of viewers. The image was created by Pablo Xavier, a 31-year-old construction worker in Chicago, and posted to a Facebook group called AI Art Universe before spreading to Reddit and other platforms.[20][21]
Newsletter writer Ryan Broderick called it "the first real mass-level AI misinformation case." While the image contained telltale signs of AI generation (such as distorted hands and overly sharp skin textures), the overall quality was convincing enough to deceive casual viewers. Shortly after the incident, Midjourney discontinued its free trial program. The company also implemented content restrictions; for example, the word "pope" was temporarily restricted, and a mid-2023 algorithm update further enhanced the platform's content filters around real public figures.[20][21]
Midjourney's use of copyrighted images as training data has become the subject of multiple high-profile lawsuits:
Midjourney has argued that its training falls within fair use parameters. In response to Disney and Universal, the company's lawyers wrote that "copyright law does not confer absolute control over the use of copyrighted works" and that "fair use safeguards public interests in the free flow of ideas and information."[70] These lawsuits are part of a broader wave of more than 70 copyright cases filed against AI companies as of 2025, and their outcomes may set significant precedent for the entire generative AI industry.[66]
Beyond the Pope Francis incident, Midjourney's improving photorealism has raised ongoing concerns about its potential for deepfakes and visual misinformation. A mid-2023 algorithm update further sharpened the platform's depiction of real public figures, including a series of fake images of a Pentagon attack that briefly moved markets.[21][71] Midjourney maintains content moderation policies and employs a small team of moderators, and the platform prohibits adult content, gore, sexualized images of real people, and offensive or inflammatory depictions of public figures. Some prompts (such as the names of named politicians during election periods, and sensitive religious terms) are blocked or filtered automatically.[72]
Midjourney V7 launched in alpha on April 3, 2025, and became the default model on June 17, 2025.[24][25] Holz described V7 publicly as "our most intelligent, aesthetically pleasing, and coherent model to date." V7 was a complete rebuild rather than an incremental update, and its main innovations were:
--oref), a unified character and object reference system controlled with --ow (omni weight, range 1 to 1000, default 100). Unlike the V6 --cref parameter it replaces, Omni Reference handles external photos well and works across poses, outfits, and environments.[46][47]Midjourney entered the video space on June 18, 2025, with V1 Video. The image-to-video workflow integrates with the existing image pipeline, allowing users to animate stills into 5-second clips that can be extended to roughly 21 seconds total. Each job produces four clips at approximately 8x the GPU cost of an image job, which Holz framed as roughly the cost of one image per second of video and "over 25 times cheaper than what the market has shipped before."[26][27]
On July 25, 2025, looping clips, custom start and end frames, and Discord-side video generation rolled out.[28] On July 30, 2025, Midjourney launched midjourney.tv, a 24/7 livestream of trending V1 Video clips along with personalized feeds.[29][30]
On August 22, 2025, Meta Chief AI Officer Alexandr Wang announced a partnership with Midjourney to license its image and video generation technology. The deal brings Midjourney's "aesthetic technology" into Meta's ecosystem, with plans to integrate AI-generated tools into Facebook, Instagram, and WhatsApp. Financial terms were not disclosed.[31][32] Midjourney confirmed that it would continue operating independently. Holz wrote on X that Midjourney remains "a community-backed research lab" with "no investors," and that the Meta deal does not involve equity.[31][32]
On March 17, 2026, Midjourney released the V8 alpha at alpha.midjourney.com. V8 introduces native 2K HD rendering via --hd (eliminating the need for separate upscaling), a new --q 4 quality mode, generation speeds roughly 5x faster than V7, and substantially improved prompt adherence, especially for text inside quotation marks. The alpha also brought a new grid view, sidebar settings, an improved conversation mode, and updated style reference and moodboard algorithms (--sv 7).[33][58]
On April 30, 2026, Midjourney released V8.1 on midjourney.com and Discord. V8.1 is the fastest model Midjourney has shipped, with standard jobs running 4 to 5x faster than earlier versions and HD mode 3x faster and 3x cheaper than V8.0, prompting Midjourney to make HD the default. V8.1 brought back image weights and image prompts, introduced an automatic Prompt Shortener for over-long prompts, and tightened moodboard and --sref stability. V8.1 requires users to unlock their Global V7/V8 personalization profile.[34][75]
In a March 2026 interview, Holz disclosed that Midjourney was exploring a small hardware project, informally called "Orb," which would be led by Holz and Midjourney head of hardware Ahmad Abbas, both former Leap Motion engineers. Holz indicated that the program could conceivably require Midjourney's first outside capital, a notable departure for a company that has rejected venture capital for five years.[13][76]
Midjourney sits within a broader ecosystem of text-to-image models and AI art tools. Adjacent technologies include the DDIM (Denoising Diffusion Implicit Models) sampler family, SDXL and other open-weight diffusion checkpoints, the Diffusion Transformer (DiT) architecture used by newer rivals, and the MMDiT backbone that underpins Stable Diffusion 3. In video, Midjourney V1 competes with Sora, Veo 3, Kling, Pika, Runway Gen-4, and Runway Aleph. In open-weight image generation, Midjourney is often compared with FLUX.1, FLUX.2, Stable Diffusion 3.5, and SDXL. Among closed services, the most direct comparisons are DALL-E 3, gpt-image-1, Imagen 3, Imagen 4, Adobe Firefly, and Ideogram 3.0.
After V8.1 reached midjourney.com and Discord on April 30, Midjourney spent May refining the model and the web experience rather than shipping a new version. The V8.1 launch post noted that standard-definition mode had been made a temporary default during a server transition, with users able to switch back to HD through the settings panel or by adding --hd, and confirmed that community rating feedback was feeding the next model, V8.2.[77]
Earlier in the V8 cycle the company also widened access and gathered ranking data. On March 21, 2026, Relax mode was turned on for V8 across Standard, Pro, and Mega subscribers, supporting most commands.[78] On April 27, 2026, ahead of what it called "major aesthetic updates," Midjourney ran a ranking session at full 2K resolution (the first time it had ranked raw images at that resolution) to tune V8.1 and V8.2 toward more natively HD output, inviting subscribers to compare image pairs at midjourney.com/rank-v8-1.[79]
On May 27, 2026, Midjourney pushed a batch of web platform updates. Conversational mode, including voice input, gained access to a session's Image Prompts, Style References, sidebar settings, and recent jobs, and tray images now persist across voice submissions. A new "Rerun as HD" button lets users re-render any V8.1 image generated in standard definition at HD. The update also added a hidden-item count to the Create and Organize folder views, grouped Profiles, Moodboards, and Liked Styles in the mobile settings menu, and enabled search for signed-in members without an active subscription.[80]
As of early June 2026, V8.1 remained the current default image model, with V8.2 in active development and no V2 video model or broadly available public API yet released.[77][80]