On this page
Midjourney in 2026: Models, References, Editing, and Practical Workflows
(This guide was updated in 2026 to reflect the latest technical changes).What is current in 2026?
The current default Midjourney version documented by Midjourney is V8.2. V8.2 became the default on July 24, 2026, and includes the Edit Model. V8.1 remains relevant because it is also compatible with the Edit Model. V7 is an older version that remains documented and is the version associated with Omni Reference.
V8.0 Alpha is not a current selectable version. The official version page says that V8.0 Alpha was available on the Alpha website until July 24, 2026. Images made with it can remain in a gallery, but the version is no longer available for new use. This is why a current guide should not describe V8.0 as the present model.
V7 was released in April 2025 and was the default version until June 9, 2026. It is useful as context when a documented feature is V7-specific, but this is not a V7-only guide. If the result depends on a particular model, include the version in the prompt or confirm the version selected in Settings.
Choose the reference type before writing the prompt.
Midjourney has several kinds of image input. They do different jobs, and they should not be treated as interchangeable. Image prompts influence content, composition, and color. Style references influence the visual treatment. The Edit Model is the documented path for modifying an image or combining reference images in V8.1 and V8. 2. Omni Reference remains available with V7.
References are sources of guidance for a new output. They are not a command to reproduce an uploaded image exactly. A reference can be interpreted differently by different models, prompts, weights, and aspect ratios. Describe the required subject, setting, materials, viewpoint, and constraints in text instead of assuming that the reference will carry every detail.
Image Prompts
An image prompt is an image attached to a prompt to influence the content, composition, or colors of a new creation. The official documentation describes three common patterns: one image with text, multiple images without text, or multiple images with text. When text is included, describe the desired final scene rather than writing instructions to copy or modify the reference.
In Discord, an image URL can be placed at the beginning of the prompt. On the website, use the image panel and attach the image as an image prompt. The file must be in a supported format such as PNG, GIF, WEBP, JPG, or JPEG. Cropping the source to a similar aspect ratio can make the intended composition easier to interpret, but it does not lock the layout.
--iw controls the influence of an image prompt. The documented range varies by version: the Image Prompts page lists a default of 1, with ranges of 0–3 for V8.1 and V7 and 0–2 for Niji 7. Confirm the range for the version and mode you are using.
Style References
A style reference communicates a visual direction, such as color relationships, medium, texture, or lighting. It does not copy the objects or people in the reference. Use a style reference when the subject should be newly described in the text prompt but the visual treatment should be related to the reference.
Use --sref the following image URL in Discord, or place an uploaded image in the Style Reference area on the website. A text prompt is required with a style reference. The --sw parameter controls style influence; the documented default is 100, and the documented range is 0–1000. Style Reference behavior can differ between V6 and V7, and the documentation also describes Style Reference versions through --sv.
Keep the text focused on the subject and scene. If the reference already supplies the visual direction, adding many competing style labels can make the intended look less clear. Style codes and uploaded style images are different features: a style code cannot be created from an uploaded image.
Omni Reference for V7
Omni Reference places a character, object, vehicle, or non-human creature from one reference image into a new V7 creation. It accepts one image and requires a text prompt. In Discord, use --oref the following image URL to control the amount of detail drawn from the reference; the documented default is 100, with a range of 1–1000.
Omni Reference is compatible with V7, not V8.X. The official documentation directs V8.X users to the Edit Model instead. Omni Reference also has compatibility limits: its results are not currently compatible with Vary Region, Pan, or Zoom Out, and an Omni Reference result may need to be opened in the Editor after removing the reference. It is also not compatible with some modes and parameters listed in the official documentation.
Details such as logos, freckles, and clothing marks may not match exactly. A clear prompt remains important, even when the reference contains the subject you want to use. You must have the necessary rights to upload the reference and should not use images of people in abusive, deceptive, or otherwise prohibited ways.
The Edit Model for V8.X
The Edit Model is documented for V8.1 and V8.2. It supports three related tasks: making changes to an existing image using written instructions, creating a new image from up to four reference images, and using the Editor for targeted inpainting or outpainting.
In Discord, it --edit is used with image URLs. On the website, attach images to the prompt or open an existing image and choose an Edit action. The model automatically tries to match the aspect ratio of the first uploaded image. Add --ar when a different ratio is required, because the account's default aspect ratio does not automatically apply in this workflow.
The Edit Model can accept descriptive prompts as well as direct instructions about a change. It can combine up to four reference images, but combining references does not guarantee that every attribute will stay separate or consistent. The documentation notes that the Edit Model is not currently compatible with --tile, and Edit Model results are not currently compatible with, Remix.
The Editor
The web editor works with Midjourney images and, through the Edit tab, personal images. It provides tools for moving and resizing the canvas, changing the aspect ratio, erasing and restoring areas, Smart Select, layers, inpainting, outpainting, and retexturing through the Edit Model in V8.X. The exact controls available can depend on whether the lighter Editor view or the full Edit tab is open.
Use a selection or erased area when only one part of an image needs to change. Use outpainting when the canvas needs to extend beyond its original frame. Layers can add other images to a composition, but visible areas are treated differently from transparent areas when an edit is submitted. Review the result instead of assuming that an unchanged area will remain mathematically identical in every workflow.
The editor can export an image by upscaling it to the gallery or downloading it. The documentation also describes an option to save the original generation and, where applicable, a transparent PNG of erased areas. For V8.X Editor creations, the documentation says that results appear on the Create and Organize pages and may be visible to other members unless Stealth mode is enabled.
Parameters that are useful to understand
Parameters go at the end of the prompt. Leave a space before the dashes, do not add punctuation to the parameter value, and do not place ordinary prompt text after the parameters. Compatibility changes by model, mode, and feature, so treat the following as a map of documented controls rather than a promise that every combination will work.
--archanges the aspect ratio. It is useful for a portrait, landscape, square, or other intended format.--iwchanges the influence of an image prompt. Its allowed range depends on the version.--srefattaches an uploaded style reference or a style reference code.--swchanges its influence.--orefattaches one Omni Reference image for V7 and--owchanges its reference weight. For V8.X, use the documented Edit Model path instead.--editinvokes the Edit Model with image URLs in the documented Discord workflow. The Edit Model is documented for V8.1 and V8. 2.--seedsupplies a seed value for controlled comparisons and experiments. A seed is not a guarantee of identical output across versions, settings, or changed prompts.--stylizechanges how much the system's default aesthetic influences the result. A higher value can compete with literal prompt details.--chaoschanges the variation between the generated options. It does not select a particular composition or guarantee novelty.--nostates elements to exclude. It is a request, not a reliable object-removal or safety guarantee, and the wording can be interpreted in context.
--cref and --cw should be treated as legacy or feature-specific controls, not as a current general-purpose method for locking a character or keeping a room identical. The official parameter and legacy-feature documentation determine when they apply. Do not treat a recipe --cref room --cw 80 as a documented way to maintain room continuity.
Practical examples
These examples are teaching prompts, not claims that a particular output is guaranteed. Replace the placeholders with images you are allowed to use. Add a version only when the workflow requires one and confirm that the selected version supports the attached references.
Product image with an image prompt
studio product photograph of a matte black insulated bottle on a pale stone surface, soft side light, clear space around the object, no visible brand name, clean commercial composition --ar 4:5 --iw 1Attach a reference that you have permission to use as an image prompt. The text describes the desired product scene; it does not instruct Midjourney to reproduce the source photograph.
Thumbnail concept
wide educational video thumbnail about repairing a vintage camera, one camera centered on a workbench, strong separation between subject and background, readable empty area on the left for later title design, no fake words or logos --ar 16:9 --chaos 10Keep final text, logos, and legal marks in a separate design pass unless they have been checked carefully. An image model can produce letter-like shapes that are not usable wording.
Concept interior
concept interior of a compact reading room, oak shelving, linen chair, indirect daylight, visible floor and ceiling junctions, calm neutral palette, straight architectural lines, no people, no text --ar 3:2 --stylize 100This is a visual concept, not a measured plan. Use drawings, specifications, and a qualified professional for dimensions and construction decisions.
Style Reference example
editorial illustration of a cyclist waiting at a quiet train platform, limited two-color palette, crisp ink contours, broad paper texture, simple background shapes --sref IMAGE_URL --sw 150 --ar 4:5Replace IMAGE_URL with a style image you have the right to use. The Style Reference should influence the visual treatment; the prompt supplies the new subject and setting.
A workflow that keeps decisions traceable
- Define the goal. Write down the intended use, audience, format, subject, and elements that must not appear. Separate visual exploration from facts that must be exact.
- Choose a lawful source. Use an image, logo, document, or photograph only when you have the necessary permission or other lawful basis. Do not upload a client's private file, a person's photo, or confidential material without authorization.
- Write the prompt. Describe the subject, setting, composition, materials, lighting, viewpoint, and exclusions. State what must be added in a later design or editing tool.
- Select the model and reference type. Use V8.2 by default unless a documented compatibility need points to V8.1 or V7. Use Image Prompts for content influence, Style References for visual treatment, the Edit Model for V8.X image editing or multiple references, and Omni Reference only for its V7 workflow.
- Generate alternatives. Change one meaningful variable at a time when comparing results. A seed can help organize a comparison, but it does not make different versions or prompts identical.
- Edit deliberately. Open the Editor for a targeted replacement, a canvas extension, an aspect-ratio change, or a layer. Keep the edited area small when the rest of the image should remain a visual anchor.
- Inspect the result. Check all text, symbols, logos, faces, hands, object counts, measurements, perspective, edges, and continuity between related images. Zoom in; do not rely on a thumbnail preview.
- Record the version. Save the prompt, parameters, reference type, source permission, date, model, and selected output. This makes later review possible when the service changes.
- Export and review outside Midjourney. Confirm the file format, dimensions, color needs, accessibility, and any required human design or fact-checking before publishing.
Limits you should plan around
Midjourney is an image-generation and editing service, not CAD or BIM software. It does not establish scale, dimensions, structural tolerances, material specifications, code compliance, or construction feasibility. An interior concept can suggest a mood, but it is not a measured drawing or a substitute for architectural, engineering, or other qualified review.
References can guide a new image, but they do not guarantee the same person, object, room, camera position, lens, lighting, or geometry across a series. The service does not guarantee correct text, trademarks, logos, hands, faces, object counts, or measurements. Inspect every result that communicates factual or commercial information.
Midjourney's Terms say that the service is provided without promises or guarantees about assets. Midjourney does not provide legal advice about copyright, trademark, publicity rights, or other rights. Ownership and permitted use can depend on the Terms, third-party rights, local law, and the way an image was created or reused. Treat an output as something to review, not as proof that a publication is legally cleared.
Plans, privacy, and commercial use
The official plan page lists four subscription tiers: Basic, Standard, Pro, and Mega. Standard, Pro, and Mega include unlimited image generations in Relax Mode according to the comparison page. Stealth Mode, which is intended to keep images and videos private on the Midjourney website, is available on Pro and Mega. Plan availability and features can change; this article does not reproduce prices.
The plan page states that a company making more than 1,000,000 USD in gross revenue per year must purchase Pro or Mega to use its images commercially. That condition is not a copyright guarantee and does not remove the need to check third-party rights or applicable law. The commercial-use page says that users own images and videos they create to the extent described in the Terms, while an upscaled image created by another user remains associated with the original creator.
The Terms say that content is publicly viewable and remixable by default when posted in a public setting. Stealth mode changes the website visibility commitment described by Midjourney, but content in a shared or open Discord space can still be visible to people in that space. Privacy is therefore a workflow decision, not an assumption about a private browser tab.
The Privacy Policy says that Midjourney may receive prompts, images, videos, documents, messages, account details, technical information, usage data, and other information supplied through the service. It also describes uses, service providers, legal disclosures, retention, and possible transfers to locations where data-protection rules may differ. Read the current Privacy Policy for the details that apply to your situation.
Do not upload client files, confidential documents, personal records, private photographs, or identifiable people without the required permission and a clear reason to use the material. You are responsible for the content you provide and for having the necessary rights to provide it.
Troubleshooting
The reference is ignored or the result is too different
Confirm that the image is attached to the intended reference type. Add a direct text description of the subject and composition. Adjust one weight at a time, and verify that the selected version supports the parameter. An Image Prompt or Style Reference guides a new output; neither is an exact-copy control.
The Edit Model changes the whole composition
Use a narrower selection in the Editor and describe only the required change. If you uploaded a source image, check the first-image aspect-ratio behavior and add --ar when appropriate. Keep in mind that the Edit Model is documented for V8.1 and V8.2 and has its own compatibility limits.
An Omni Reference result cannot be edited as expected
Check that the job is a V7 job. The official documentation says that Omni Reference results are not compatible with Vary Region, Pan, or Zoom Out. Open the image in the Edit tab and remove the Omni Reference and --ow before submitting an edit where the Editor requires that change.
Text or a logo is wrong
Regenerate only if the image concept is still useful, but plan to add exact wording and brand marks in a controlled design tool. Inspect small type, punctuation, symbols, and repeated letters. Do not publish an invented mark as if it were an approved logo.
The output is not private
Review where the job was created and whether Stealth mode was enabled. The Terms describe public visibility and remixing as the default in public settings. A shared Discord room can expose content to room participants even when other privacy settings are enabled.
Frequently asked questions
Should I follow a V7-only tutorial?
Only for a feature that is documented as V7-specific. V8.2 is the current default. V7 remains useful context for Omni Reference and older workflows, but parameter compatibility and feature behavior should be checked against the current documentation.
Can I use an image from the internet as a reference?
Use it only when you have the necessary rights or permission. Midjourney's external-image guidance makes the uploader responsible for the content they provide. A public URL is not the same as permission to reuse an image.
Does a seed make a series consistent?
A seed can help organize comparisons, but it is not a guarantee of identical results. Model version, prompt wording, parameters, references, and service changes can alter the output.
Can Midjourney create a precise floor plan or BIM model?
No. It can create a visual concept, but it does not establish verified measurements, construction data, or BIM objects. Use appropriate technical software and qualified review for those needs.
Can Midjourney guarantee copyright or non-infringement?
No. Midjourney's commercial-use documentation says it cannot provide copyright guidance. Rights depend on the Terms, the inputs, third-party rights, local law, and the facts of the work. Consult a qualified lawyer for a specific legal question.
Which plan should I choose?
Compare Basic, Standard, Pro, and Mega using the current official plan page. Consider whether you need Relax Mode, Stealth Mode, or the commercial-use condition for a company above the stated revenue threshold. Do not rely on old price lists or feature tables.
Sources
- Midjourney Version
- Midjourney Image Prompts
- Midjourney Style Reference
- Midjourney Omni Reference
- Midjourney Edit Model
- Midjourney Editor
- Midjourney Parameter List
- Comparing Midjourney Plans
- Using Images and Videos Commercially
- Midjourney Terms of Service
- Midjourney Privacy Policy
Information checked against the linked official Midjourney documentation on September 28, 2026; models, features, plans, privacy settings, and compatibility may change.