CtrlK
Midjourney V7 is the seventh major image generation model developed by Midjourney Inc. It launched in alpha on April 3, 2025, and became the platform's default model on June 17, 2025. V7 introduced three headline capabilities: a personalization system turned on by default, Draft Mode for rapid low-cost generation, and Omni-Reference for consistent character and object rendering across multiple images. CEO David Holz described the underlying system as "a totally different architecture," and the model delivered notable improvements in texture quality, anatomical coherence, and prompt understanding compared to its predecessor, V6.1.
By the time V7 became the default, Midjourney reported that personalization profiles had been improved and were preferred by 85% of users. The model competed during 2025 against a crowded field that included FLUX 2 Pro, GPT Image 1, and Imagen 4, each of which held advantages in specific domains. Midjourney V8, rebuilt on a new GPU-native codebase, entered alpha testing on March 17, 2026, and V8.1 followed on April 14, 2026.
Midjourney Inc. has released its image models in numbered versions since the platform's launch in July 2022. The first four versions iterated on diffusion-based approaches, with each generation improving stylistic quality and prompt responsiveness. Version 5 (March 2023) moved toward higher photorealism and introduced the --ar aspect ratio parameter. Version 5.1 and 5.2 followed in May and June 2023, focusing on detail sharpness and the --style raw option for less interpretive outputs.
Version 6, released on December 20, 2023, made significantly longer text prompts practical and improved multi-subject scene coherence. It became the default model on February 14, 2024. Version 6.1 arrived on July 30, 2024, adding more coherent limb rendering, more precise fine details (eyes, small faces, distant hands), and a 25% speed improvement. A major product milestone in August 2024 was the launch of a dedicated web interface at midjourney.com, which supplemented the existing Discord-based workflow. V6.1 remained the default until V7 took over in mid-2025.
The roughly nine months between V6.1 (July 2024) and V7 (April 2025) was Midjourney's longest gap between major model releases. During that period, the company released a series of platform features rather than model upgrades: editor tools, video generation previews, and expanded web app functionality. The long interval raised user expectations and, when V7 launched, contributed to divided reactions, with some users feeling the improvements were incremental and others praising the new tooling around personalization and speed.
Midjourney released V7 in alpha around midnight ET on April 3, 2025. The release went to all subscribers simultaneously rather than through a staged rollout. At launch, V7 was available in two speed modes: Turbo (2x GPU cost, highest speed) and Relax (no fast-hour consumption, available on Standard plans and above). Standard mode was described as coming soon and required additional optimization before activation. The web interface updated to show V7 as a selectable option in the Version dropdown, and Discord users could access it with the --v 7 parameter.
At alpha launch, several established features from V6.1 were not yet available in V7. Image upscaling, retexturing, zoom out, and vary region fell back to V6 rendering. Remix mode was also absent. Midjourney projected these would arrive within approximately two months, and the company committed to releasing new features every one to two weeks for the following 60 days. Style reference codes generated in V6.1 remained compatible with V7 prompts.
Following the alpha period and the subsequent releases of Omni-Reference and improved style references, Midjourney made V7 the default model on June 17, 2025. Users who preferred V6.1 could revert by selecting it from the web settings dropdown or by appending --v 6.1 to any prompt in Discord.
Midjourney has not published a technical paper describing V7's architecture in detail, which is consistent with the company's historically closed approach. Holz characterized V7 as "a totally different architecture" with substantially better text prompt comprehension. The observable behavioral changes suggest the model handles multi-element compositions differently than V6.1, resolving spatial relationships and maintaining subject-background coherence with greater consistency.
Third-party comparisons run after the release identified several clear improvements:
In action photography prompts, V7 produced results described by reviewers at Gold Penguin as "slightly gritty" in a way that matched actual sports photography, whereas V6.1 often produced "exaggerated motion blur" that flagged images as AI-generated. Water physics in landscape prompts showed more natural wave patterns. Fabric texture rendering improved to the point where individual threads in knits were sometimes discernible, compared to the blurred textures V6.1 produced.
Anatomical accuracy across body types improved by an estimated 40% according to aggregated reviewer assessments, with hands and fingers showing the most visible gains. Prompt adherence improved approximately 35%, meaning V7 followed detailed multi-element descriptions more reliably. Text rendering in images, historically a weakness across all Midjourney versions, improved only about 15% and remained unreliable for anything beyond short phrases.
Areas where V7 was not clearly superior included product photography, where lighting accuracy in some reviewers' tests actually regressed slightly compared to V6.1. Critics at Decrypt characterized V7 overall as "a welcoming upgrade" but noted it was "more of a model evolution, rather than the revolution we're seeing in this new wave of image generators." Some users specifically observed that V6.1 handled photographic realism with more convincing lighting physics, while V7 excelled more at stylized and painterly outputs.
V7 includes a Raw mode option (accessible via --style raw) that removes the default Midjourney aesthetic processing and produces outputs closer to the literal prompt. This is useful for users working on photorealistic mockups or product visualization who find V7's default stylistic interpretation too pronounced. Raw mode in V7 is not the same as the --style raw behavior in V6.1 due to the architectural differences, and users migrating from V6.1 workflows typically needed to re-test their prompts.
Personalization is the feature Midjourney most prominently positioned as V7's defining capability. It is the first Midjourney model to have personalization enabled by default.
Personalization builds a preference profile by having the user rate images. To unlock the Global Personalization Profile, users visit the Personalize page on midjourney.com and select images they find most appealing from a grid of options. The process takes approximately five minutes and requires selecting enough images to fill a progress bar, which corresponds to rating roughly 200 image pairs. The system learns from these ratings which visual qualities the user prefers: vibrant versus muted palettes, realistic versus stylized rendering, minimalist versus maximalist composition, and similar aesthetic dimensions.
Once unlocked, a personalization icon appears next to the Imagine bar in the web interface. When active, the user's personalization profile is automatically appended to every prompt. The --p parameter in Discord commands applies the same profile. Users can toggle personalization on or off at any time without rebuilding the profile.
Multiple personalization profiles are possible. Each profile generates a unique alphanumeric code, and users can append different profile codes to specific prompts to switch between learned aesthetics. This allows, for example, one profile trained toward editorial photography and another toward illustration, with the user selecting the appropriate one per project.
The degree to which the personalization profile influences output is adjustable via a slider in the web interface. In Discord, the --p parameter accepts a numeric argument controlling profile weight. Setting personalization to a lower value applies the learned preferences subtly, while higher values tilt outputs more strongly toward the learned aesthetic. Midjourney does not publish the exact scaling formula.
The personalization system received broadly positive user feedback. The company's June 2025 announcement that V7 had become the default cited internal data showing personalization profiles were preferred by 85% of users. Some power users noted that the system's influence was difficult to predict precisely, and that combining personalization with specific style references (--sref) produced different results than either parameter alone. Others found value in creating profiles for clients with defined brand aesthetics.
Draft Mode is a new generation option introduced with V7 that produces images at 10 times the speed of standard V7 rendering at half the GPU cost. Draft outputs are lower in resolution and detail than standard renders, but they maintain consistent color palette, subject placement, and compositional structure, making them useful as a pre-visualization step before committing to full-quality generation.
In the web interface, Draft Mode activates a conversational UI. Users can describe modifications verbally or in text (for example, "swap the cat with an owl" or "make it night time"), and the system adjusts the prompt and starts a new job automatically. Voice mode is supported in this interface, allowing users to iterate on images through spoken instructions without retyping prompts. In Discord, Draft Mode is triggered with the --draft flag.
Generating a draft image does not prevent upgrading it. A single click in the web interface sends the draft to full standard rendering, which produces a higher-quality version from the same prompt. This two-step workflow reduces exploratory costs significantly: users can test ten compositional directions at the cost of roughly five standard images.
Draft Mode is not compatible with Omni-Reference (--oref). It also cannot be combined with Fast Mode or the --q 4 quality setting. Draft Mode works with style references and personalization. The lower resolution of Draft outputs means they are not suitable for final deliverables and should be treated as concept exploration only.
Omni-Reference (accessed via --oref in Discord or through the reference image slot in the web UI) allows users to provide a single reference image that the model uses to reproduce the visual characteristics of a character, creature, object, or product across different scenes and compositions. It was released after V7's alpha launch and was one of the features cited as finalized before Midjourney made V7 the default.
The user provides a reference image URL or uploads an image into the designated Omni-Reference slot in the web interface. Midjourney extracts shape, color, and anatomical features from the reference and attempts to reproduce those features in the generated output, even when the scene, lighting, or angle changes substantially. This differs from style references (--sref), which transfer aesthetic qualities like color palette and texture rather than specific subject identity.
The --ow (omni-weight) parameter controls how strongly the reference image influences the output, on a scale from 1 to 1000 with a default of 100. Values between 25 and 50 produce a looser stylistic transfer where only general visual similarity is maintained. Values between 100 and 300 create a balanced influence where the reference subject is recognizable but adapts naturally to new contexts. Values above 400 produce maximum fidelity and are suitable for scenarios requiring precise visual identity, such as product packaging shots from multiple angles.
Omni-Reference is only available in V7. It is not compatible with Draft Mode, Fast Mode, Conversational Mode, inpainting, outpainting, or the --q 4 quality parameter. Only one reference image per prompt is supported; multi-reference conditioning is not available in V7 (FLUX 2 supports up to ten reference images). Each Omni-Reference job consumes approximately twice the GPU time of a standard V7 generation.
Omni-Reference addresses one of the most persistent complaints about generative image models: the inability to maintain consistent character appearance across multiple outputs. Prior to V7, Midjourney's character reference (--cref) was the closest equivalent and worked primarily for human faces. Omni-Reference extends consistency support to non-human subjects, objects, creatures, and products, which makes it practical for:
In parallel with the main V7 model, Midjourney's Niji series (developed in collaboration with Spellbrush and focused on anime and Eastern illustrative aesthetics) reached version 7 on January 9, 2026. Niji 7 brought improvements in coherence for fine details such as eyes, reflections, and background elements, and the team intentionally reduced the 3D-rendered appearance of Niji 6 to emphasize traditional anime linework. Prompt understanding in Niji 7 became more literal, following specific instructions with higher fidelity than previous versions. The character reference feature (--cref) was removed from Niji 7, with the Midjourney team indicating a replacement feature was in development.
Midjourney operates on a subscription model with no free tier as of early 2023. All plans give access to V7. The four plans as of 2025 are:
| Plan | Monthly price | Annual price (per month) | Fast GPU hours | Concurrent jobs | Stealth mode |
|---|---|---|---|---|---|
| Basic | $10 | $8 | ~3.3 hours | 3 | No |
| Standard | $30 | $24 | ~15 hours | 3 | No |
| Pro | $60 | $48 | ~30 hours | 12 | Yes |
| Mega | $120 | $96 | ~60 hours | 12 | Yes |
Relax mode (slower generation with no fast-hour consumption) is available on Standard and higher plans. Stealth mode, which keeps generated images private rather than publicly visible in the community gallery, requires a Pro or Mega subscription. An annual billing option reduces the monthly cost by approximately 20%.
Within these plans, different generation types carry different GPU costs. Turbo mode costs 2x the standard rate. Draft mode costs 0.5x. Omni-Reference jobs cost approximately 2x the standard rate due to the additional reference processing. There is no free tier, and there has not been one since April 2023 when Midjourney removed trial access.
Midjourney released V8 in alpha on March 17, 2026, available on alpha.midjourney.com to all paid subscribers. V8 is the first Midjourney model built on a completely new GPU-native codebase, representing an architectural break rather than an incremental refinement of V7.
Key differences from V7 include approximately 5x faster generation, native 2K resolution through a --hd mode, and substantially better text rendering in images. Multi-element compositions that V7 would partially ignore now render with noticeably higher fidelity in V8. The model follows detailed, specific prompts more literally, which is described as a philosophical shift: V7 fills gaps with creative interpretation, while V8 prioritizes literal adherence to specifications.
V8 introduced a Style Creator tool that allows users to extract a visual style from a reference image and save it as a shareable code. Teams can distribute style codes without revealing the original reference images, which supports consistent branding across multiple contributors.
At V8's alpha launch, advanced features including --hd mode, --q 4, style references, and moodboards cost 4x the standard rate. Relax mode was not available. Draft Mode was absent from the V8 alpha, which reviewers noted as a significant workflow regression for users accustomed to the low-cost iteration loop in V7. Personalization in V8 was described as early-stage, compared to the mature system in V7.
V8's visual aesthetic shifted from V7's warm, painterly, editorially stylized outputs toward cooler, more photographically literal rendering. Reviewers described V7 as better for fantasy, concept art, and stylized creative work, while V8 suited product photography, architectural visualization, and commercial applications requiring physical accuracy.
V8.1 entered alpha on April 14, 2026, designed to retain V8.0's technical improvements while restoring the aesthetic continuity and workflow features users associated with V7. V8.1 described its visual output as "a consistent and familiar aesthetic in the spirit of V7," and moodboards and style references became notably more stable. HD mode performance improved 3x in speed and reduced 3x in cost, becoming the default resolution for V8.1 rather than an opt-in premium feature. Standard resolution in V8.1 reached the same speed as V7's Draft Mode at full quality. Midjourney announced plans to discontinue V8.0 after V8.1 stabilized, with future development focus on upscalers and then editing, inpainting, and outpainting.
Midjourney V7 launched into a competitive field. The table below summarizes the primary alternatives as of mid-2025.
| Model | Developer | Access | Strengths | Weaknesses | Pricing |
|---|---|---|---|---|---|
| Midjourney V7 | Midjourney | Subscription (web, Discord) | Artistic style, ease of use, Draft Mode, Omni-Reference | Text rendering, multi-reference, no API | $10-$120/mo |
| FLUX 2 Pro | Black Forest Labs | API, subscription | Text rendering, multi-reference (10 images), precision, API access | Less stylized aesthetic, API complexity | ~$0.03-$0.12/megapixel |
| GPT Image 1 | OpenAI | ChatGPT subscription, API | Prompt adherence, text accuracy, conversational editing | Less cinematic, subscription required for ChatGPT access | Bundled with ChatGPT Plus / API pricing |
| Imagen 4 | Google DeepMind | Gemini API, AI Studio | Product photography, specular accuracy, edge cleanliness | Less creative range, limited consumer access | API (pay per image) |
FLUX 2 Pro uses a 32-billion-parameter latent rectified-flow transformer paired with a 24-billion-parameter vision-language model and is designed as an infrastructure system for production pipelines. It excels at consistently legible typography, physically plausible photorealism, and multi-reference conditioning with up to ten simultaneous input images. FLUX 2 achieves 88-92% accuracy on multi-word text rendering, compared to approximately 52% for Midjourney V7.
Midjourney V7 produces outputs described as more "cinematic" and dramatically stylized, with richer contrast and atmosphere, and requires minimal prompt engineering to achieve visually striking results. FLUX 2 tends toward an over-smooth, idealized skin tone (described by some reviewers as "Flux shine") and is less willing to embrace highly stylized aesthetics.
For production workflows at scale, FLUX 2's per-megapixel API pricing is more predictable than Midjourney's subscription-plus-GPU-hour model. For creative exploration, mood boards, and concept art, Midjourney V7 remains the more widely used choice due to community infrastructure and ease of use.
OpenAI integrated image generation directly into GPT-4o with the GPT Image 1 update in March 2025, replacing the DALL-E 3 backend with an autoregressive architecture tied to GPT-4o's language understanding. GPT Image 1 is substantially stronger at prompt adherence for complex instructions involving counting objects, precise spatial relationships, and text rendering within images.
In comparative tests using typical user prompts, ChatGPT's image generation and Midjourney V7 trade results depending on the prompt type. For prompts requiring precise visual accuracy (specific number of objects, legible text, exact described compositions), GPT Image 1 performs better. For prompts emphasizing aesthetic quality, mood, and artistic interpretation, Midjourney V7 typically produces more visually compelling results. The Decrypt review noted that for comparable subscription costs, ChatGPT offered "superior prompt adherence and spatial awareness."
Google's Imagen 4, available through the Gemini API and Google AI Studio, targets product photography and commercial imagery. In a 2025 benchmark by Everypixel comparing product shot realism, Imagen 4 ranked first on specular accuracy (how correctly it renders reflective surfaces) and edge cleanliness (precise object separation from backgrounds). For isolated product photography with transparent backgrounds, Imagen 4 leads the field.
Midjourney V7 produces the strongest raw hero assets for editorial and conceptual use cases where aesthetic quality matters more than physical precision. Consumer access to Imagen 4 remained limited throughout 2025, with most availability through paid API tiers.
V7's combination of personalization, Draft Mode, and Omni-Reference expanded the practical range of applications compared to V6.1.
Commercial applications include product visualization using Omni-Reference to place a specific product in multiple lifestyle contexts, brand mood boards generated at Draft Mode speeds for client presentation, and character sheets for games, animation, or publishing that require consistent visual identity across multiple angles and scenarios.
Personal creative work includes fine art exploration, where personalization profiles reduce the iteration time needed to arrive at outputs matching individual aesthetic sensibilities. Sequential storytelling and webcomic production benefit from Omni-Reference for recurring characters.
Interior designers and architects used V7 for spatial concept visualization. Side-by-side testing by Mark Joseph Lape published on Medium compared V6.1 and V7 for interior design prompts and found V7 produced more refined material textures and more convincing ambient lighting in room scenes.
Marketing teams used Draft Mode as a rapid ideation layer, generating dozens of compositional directions in the time previously required for a few standard renders, then upgrading only the strongest candidates.
The initial reception to V7 was mixed. On Reddit and X (formerly Twitter), many users praised the improvement in hands, bodies, and textures. Others expressed disappointment that the nine-month wait had not produced a more dramatic upgrade. Some users specifically noted that V7 occasionally felt like "the model is trying to impress rather than interpret," producing stylized outputs that prioritized aesthetic impact over faithfully executing the user's described intent.
The Decrypt review summarized the competitive situation plainly: "Unless users are already invested in Midjourney's Discord community, the upgrade provides insufficient differentiation to attract new users from established competitors." The review pointed to Reve and ChatGPT as offering superior prompt adherence for similar or lower costs.
Reviews from practitioners in specific domains were more positive. Interior designers and photographers noted that action shots and material textures showed clear, practical improvement. The Substack publication by Gareth B. Davies cataloged multiple prompt categories and found V7 consistently better for fashion photography, landscape work, and editorial portraiture.
By June 2025, when Midjourney made V7 the default, the model had had approximately three months to accumulate user feedback, and the company's internal data (the 85% personalization preference figure) suggested that users who engaged with the new features found them valuable.
Several limitations were documented across V7's production period:
Text rendering remained unreliable for anything beyond very short phrases. Both V6.1 and V7 produced similar failure modes on longer text strings in comparative tests. FLUX 2 and GPT Image 1 both performed substantially better on this dimension.
Multi-reference conditioning was limited to a single image via Omni-Reference. Users needing to maintain consistency across multiple characters in a single scene, or to fuse elements from multiple reference images, could not do so directly. FLUX 2 supported up to ten reference images per prompt.
Midjourney V7 has no public API. Automation and production pipeline integration are prohibited under the terms of service, which limits its utility for high-volume commercial workflows compared to FLUX 2 or Imagen 4.
Images generated on the Basic and Standard plans (the majority of subscriptions) are publicly visible in the community gallery. Users requiring private outputs must subscribe to Pro or Mega. This creates a practical limitation for commercial work and client deliverables at the lower price tiers.
Editing consistency was flagged as a weakness. When users requested changes to a generated image through the web editor or natural language commands, the editor sometimes altered the overall scene composition or style rather than targeting only the specified element, causing unintended changes to subjects or backgrounds.
The Weirdness parameter (--weird), which introduces unexpected stylistic variation, was not available in V7 at launch and remained disabled during the alpha period.
Copyright litigation became a significant concern throughout 2025. In June 2025, Disney, NBCUniversal, and DreamWorks filed a major copyright infringement lawsuit alleging that Midjourney trained on their intellectual property and generates images reproducing their copyrighted characters. Warner Bros. subsequently joined with a separate complaint in September 2025, alleging willful infringement including video content. These legal challenges did not directly affect V7's functionality but created business uncertainty for enterprise users.