Midjourney

What Midjourney is

Midjourney is an AI visual creation service and the name of the laboratory developing it. Its public product focuses on generating and transforming images and animating them into clips. It supports aesthetic exploration: imagining a scene, comparing compositions and developing a visual direction from language and references.

The main output is an image or video, rather than a design document with every decision represented as an independently editable object. This makes it useful to distinguish concept development from finishing. A selected illustration can subsequently become part of a presentation, poster or audiovisual piece. People who need to change every letter, curve or component generally complete the workflow in dedicated layout, vector or video editing tools. Planning that handoff helps determine the composition and required output before generating large numbers of alternatives.

Reference documentation: Midjourney research lab.

Origins and product development

David Holz leads the independent Midjourney laboratory, which describes itself as self-funded. Its early public development centred on image generation and the Discord community in 2022. Official model history places V3 between July and November that year, with V4 arriving in November.

Development combined model changes with a more extensive web interface and editing tools. In March 2023, Google Cloud documented V4 training on TPUs and image generation using GPUs. This describes infrastructure at that time, rather than a complete account of every current model. V1 video launched on 18 June 2025, adding image animation to the creative workflow. That expansion allows a static concept to become a short exploration of movement, although generating a still and developing its motion remain different tasks with different controls and review needs.

Reference documentation: Midjourney research lab · Google Cloud partnership, 14 March 2023 · Legacy model history · V1 video launch, 18 June 2025.

Models and feature compatibility

As of 8 November 2026, documentation identifies V8.2 as the default model, released on 24 July 2026. The update focused on aesthetics, quality and personalisation. Earlier families remain relevant when interpreting existing images and tutorials, but their parameters cannot automatically be transferred to the current workflow.

Compatibility affects references, enlargement, editing and resolution. V8.1 and V8.2 support native 2K HD generation; some editing operations reduce output to SD before subsequent enlargement. Niji is a model line focused on anime and related aesthetics, with its own compatibility. For an editorial collection, record the version and settings used. Changing models partway through a series can alter its visual language even when the written description remains unchanged, so an update should trigger a review of consistency across the set.

Reference documentation: Version 8.2 release, 24 July 2026 · Current model compatibility · Parameter list.

From visual brief to selection

The web workflow starts in the Imagine bar, with creations organised in a gallery. Discord provides another route through commands. Text can describe the subject, environment, lighting and framing, while controls adjust aspects such as proportions, stylisation and variety. Raw changes the model’s aesthetic intervention; it does not turn generation into objective reproduction.

A useful workflow begins with a communication decision: what should someone understand when seeing the image? Alternatives can then vary one important factor at a time, such as camera distance or spatial arrangement. Judging every option solely by beauty makes selection difficult. Record why an image works: it leaves room for a headline, directs attention to the right element or avoids an ambiguous interpretation. That explanation turns a collection of outputs into an editorial process another person can continue, rather than requiring them to guess which image was chosen and why.

Reference documentation: Midjourney — Creating on Web · Prompt basics · Parameter list.

References, style and personalisation

References serve different purposes. Image Prompts influence content, composition and colour. Style Reference guides appearance, such as texture or lighting, rather than being designed to copy the depicted object. Personalization captures preferences through ratings and profiles. Choosing the appropriate feature avoids asking a style control to preserve identity.

To develop a series, prepare a small, coherent visual reference and explain which qualities should persist: palette, level of detail, viewing distance or illustration type. There is no need to preserve every object in the reference. If a cover includes a helmet, for example, its colour might belong to the style while its presence in every scene may be unnecessary. Separating those choices helps maintain coherence without mechanically repeating the same composition or carrying irrelevant props into a different subject.

Reference documentation: Image prompts · Style references · Personalization.

Local editing and image output

The Edit Model for V8.1 and V8.2 supports written instructions and up to four references, replacing earlier features such as Omni Reference and Character Reference in this workflow. The editor combines area selection, canvas expansion and layers for generated or external images. Download options include the generation or a PNG showing the current edit with transparency.

Local editing can correct one part without always starting again, but the whole image needs another review. Changing an object may alter shadows, size relationships or continuity with the background. Compare before and after at publication size and retain both files. A mask organises the intervention; it does not mean every object is available as a vector component, or that an original photograph remains unaltered. This distinction matters when the final asset must be revised repeatedly by a separate production team.

Reference documentation: Edit Model · Image editor.

Animation and video

Video generation starts from an image. Clips begin at five seconds and can be extended four times to a maximum of 21 seconds. Options include low or high motion, automatic or guided extensions, and an end frame or loop. Output can be downloaded as MP4 or GIF. Not every reference type used for still images carries over to video.

Movement introduces questions a cover image does not face: what stays stable, what changes and how does the action finish? A slow camera movement may introduce a scene; a complex manual action requires checking continuity of hands, tools and contact. For a longer piece, plan multiple shots and subsequent editing. Chaining extensions does not itself replace a script, narrative continuity or sound editing. The model’s launch announcement also acknowledged that higher motion can produce distortions, making motion choice a creative and practical decision.

Reference documentation: Video guide · V1 video launch, 18 June 2025.

Variation, seeds and working records

A seed fixes a generation starting point and helps compare changes. It is not a file containing a complete character or style. V8 documentation describes high reproducibility, but also explains that model version, text, settings and processing mode affect the result.

For a reusable project, retain the approved asset, description, authorised references, version, settings and selection rationale. When returning months later, that package is more useful than an isolated seed. It also distinguishes a correction to the original from a fresh interpretation. In a cover collection, for example, consistency across the series should be checked after changing the format of one piece. Keeping the actual approved output is essential because a record of generation settings serves a different purpose from the deliverable itself.

Reference documentation: Seeds.

Access, plans and integration

The reviewed offering distinguishes Basic, Standard, Pro and Mega. Plans vary GPU time, concurrency, Relax access and other capabilities. Relax image generation starts with Standard; SD video in Relax and Stealth require Pro or Mega. HD video is available from Standard in Fast mode.

Fast uses subscribed processing time; Relax uses a queue, while Turbo prioritises speed with higher consumption. Budgeting should include repetitions, extensions and discarded versions. A subscription does not provide unlimited capacity in every mode. Documentation generally prohibits automated interactions and states that no ordinary public API is provided, apart from expressly granted exceptions. A third-party integration should therefore not be assumed official simply because it uses Midjourney’s name. Assess the supported access route before building an organisational workflow around it.

Reference documentation: Subscription plans · GPU speed · Community Guidelines.

Rights, visibility and input material

The terms distinguish contractual rights over creations from third-party rights and protection under applicable law. Companies exceeding one million US dollars in annual gross revenue require Pro or Mega for commercial use under the specified terms. Users also grant the provider a licence over submitted and produced content.

This makes it necessary to decide which references may be uploaded before doing so. Possessing a photograph does not establish permission for every transformation or distribution. Visibility is a separate issue: Stealth controls exposure on the website, but does not automatically hide work made in a shared channel. For restricted material, review the working space and terms alongside image settings. The service’s commercial licence likewise does not guarantee that a particular output can be registered as a work of authorship. These are distinct decisions in an editorial approval process.

Reference documentation: Terms of Service, 27 May 2026 · Using Images and Videos Commercially · Midjourney — Stealth Mode.

Public uses, reception and documented problems

Midjourney’s adoption has brought together artistic experimentation, illustration and debate about authorship. Jason Allen’s Théâtre D’opéra Spatial was the subject of a US Copyright Office decision in September 2023: the application did not disclaim AI-generated material. This was a particular registration case, rather than a universal rule about every assisted creation.

The synthetic picture of Pope Francis in a puffer jacket, circulated in 2023, demonstrated how a convincing illustration could be mistaken for a photograph. In another field, Frontiers retracted an article in February 2024 because AI-generated figures did not meet its standards. These cases illustrate different problems: an image’s provenance and the rigour of scientific representation.

Rest of World’s 2023 investigation examined 3,000 Midjourney images and found stereotypical national representations. Those findings belong to that study and period; they are not a measurement of V8.2. For a current campaign, review who appears, which jobs are assigned to different people and which forms of diversity are absent, alongside the graphic quality of the output. An attractive set can still communicate an unnecessarily narrow view of its audience.

Reference documentation: Copyright Office: Théâtre D’opéra Spatial, 5 September 2023 · Smithsonian: synthetic Pope image, March 2023 · Frontiers retraction, 16 February 2024 · Rest of World image stereotypes investigation, 2023.

Litigation and training controversy

Litigation has concerned the use of works to train models and outputs reproducing protected material. Artists filed an action in 2023. Disney and Universal sued in June 2025, followed by Warner Bros. in September that year. Midjourney denied the allegations and defended training as transformative use. In July 2026, EVOX brought another action concerning vehicle photographs.

The reviewed sources document allegations, responses and proceedings; this profile does not present them as final judgments. Training, requested content and subsequent distribution should also be distinguished. These questions are related, but not interchangeable. For an editorial team, the practical implication is to retain the provenance and permissions of references, avoid equating technical capability with authorisation, and review outputs intended for commercial campaigns individually. A general description of an ongoing dispute cannot settle the rights attached to a particular proposed image.

Reference documentation: Artists copyright action, January 2023 · Disney and Universal action, June 2025 · Warner Bros. action and Midjourney response · Disney case public docket reproduced by Justia · EVOX complaint, July 2026.

Practical applications in OSH

In occupational safety and health, it can be used to explore campaign concepts, covers or supporting illustrations. The proposed application is editorial: visualising a general message before the prevention team selects and reviews materials for workers. It should not be assumed that an attractive output accurately represents a particular workplace or task.

It can provide visual metaphors to open a discussion, cover illustrations or campaign concepts about attention and communication. These uses need not pretend to document a real facility. A conceptual composition may communicate more honestly and clearly than a synthetic photograph with apparently precise details.

When teaching a specific operation, accurate tools, protective measures and movements become essential. If they cannot be checked, use a reviewed representation or authentic record instead of assuming that the model knows the procedure.

Worked example and result checks

Example: explore an illustration about shared attention in a workplace. The team compares compositions and keeps only those whose elements do not contradict the site’s safe practices, then adds reviewed wording and makes the illustrative nature of the image clear.

1. Write the campaign message and define the audience. Specify what someone should remember after seeing the image.
2. Describe a conceptual scene without names, logos or worker photographs. Separate essential and decorative elements.
3. Compare variants using the same questions: what action does each suggest, and what mistaken interpretation might it encourage?
4. Inspect the background, equipment and postures as well. Reject contradictory messages even when the overall composition is appealing.
5. Add approved wording in an editor and test it at final size. Identify the scene as an illustration and retain the chosen assets.

This proposed process helps choose a creative direction; it does not certify a scene’s safety or a campaign’s effectiveness.

Limitations and conditions of use

Aesthetic quality can hide errors in hands, tools, protective equipment or spatial relationships. Do not use these images as plans, inspection evidence or technical demonstrations without verification. For operational instructions, prefer photographs or diagrams faithful to the actual equipment and procedure.

Review privacy before generation. Stealth controls visibility on the Midjourney website, but work created in a shared channel may remain visible to its participants. Do not upload confidential material assuming that a feature name resolves every access issue.

A photorealistic appearance should not be used to present a hypothetical accident as an actual event. Nor should generated imagery fabricate evidence of participation or conduct by real people. The caption and surrounding text must preserve the distinction between illustration and documentation.

Reference documentation: Midjourney — Creating on Web · Midjourney — Stealth Mode.

When to choose it and how to complement it

Midjourney fits aesthetic exploration of a concept. For a family of graphics subsequently adjusted through points and curves, consider the vector workflow of Recraft. To turn a chosen illustration into campaign material, Canva supports final composition.

Assess selection and correction time as well as generation speed. A quickly generated image that requires essential elements to be rebuilt may be inefficient. Include the eventual reviewer in the trial so acceptance criteria reflect the intended use.

Frequently asked questions

Can it accurately depict a particular machine?

Do not assume so. Every detail needed to understand its use requires checking against documentation or authorised references.

Does Stealth make every publication private?

No. Shared-space visibility still matters; check where generation occurs and who can access it.

What is a more suitable campaign use?

Illustrating a clearly identified concept while keeping verifiable instructions and data separate from the generated image.