Let me be honest with you: the original Midjourney V7 launch in April 2025 was a bit of a mess. CEO David Holz called it “a totally different architecture” — and he wasn’t wrong. But several core features weren’t ready at launch, and users who brought their V6 workflows over verbatim got unpredictable results. By June 2025, when V7 became the default model, most gaps had been filled. What you have now is genuinely a different generation of tool.

This guide covers what actually changed, which new features are worth your time, and the specific workflow adjustments that separate sharp V7 outputs from forgettable ones. No filler frameworks. Just what works.

Midjourney V7 isn’t V6 with a new coat of paint. According to the version docs, “text and image prompts are handled with stunning precision, while image quality shines with richer textures and more coherent details — especially in bodies, hands, and objects.” That last part matters. V6’s struggle with anatomy was real. V7 doesn’t fully solve it, but the failure rate on hands is dramatically lower.

The architecture change has a side effect worth knowing: V7 prioritizes vibe and aesthetic over literal instruction. If you’ve been writing long, hyper-literal prompts — “a woman with shoulder-length auburn hair, wearing a navy blazer, seated at a mahogany desk with a laptop, afternoon light from the left” — you may find V7 interprets the mood rather than the specifics. Shorter, evocative prompts often land better now. Not always. But often enough to adjust your defaults.

Three things are genuinely new and worth spending time on: Draft Mode, Omni Reference, and the personalization engine. The rest — stylization parameters, aspect ratios, remixing — work similarly to V6 with refinements. Let’s take each new feature in turn.

This is the feature most people skip because the name implies low quality. Don’t. Draft Mode runs at roughly 10× the speed of the standard model and costs about half as much — dropping the effective cost per image from around $0.03 to $0.015. It’s not low quality. It’s low detail. For anything where you’re solving a composition problem — layout, framing, the rough relationship between elements — Draft Mode is the right tool.

Think of it the way a painter thinks about rough sketches before committing to canvas. You wouldn’t spend three hours on a detailed study to discover the composition doesn’t work. Use Draft Mode to kill bad ideas fast, find the seed you like, then upscale in standard mode.

Practical workflow: draft → find the seed you want to iterate from → copy that seed → run the final in standard mode. You’ll spend a fraction of your fast GPU hours on dead ends.

V6 had Character Reference (--cref), which was useful but inconsistent — change the lighting, the face would drift. V7 appears to understand the underlying 3D structure of a face better, which means character references hold through varied angles and lighting conditions more reliably than before.

Omni Reference goes further. It lets you anchor not just a character but any specific object — a logo, a product, a vehicle — so it stays consistent across multiple outputs. Use --ow to control how strongly the reference influences the generation. Higher values = stronger lock on the reference, less creative latitude. Lower values = more interpretation, more variation. The sweet spot depends on how distinctive your reference is.

One important limitation: text rendering improved about 15% over V6, but accuracy for complex sentences still sits around 30%. If your use case requires readable text inside the image — signage, labels, readable titles — V7 is still not the right tool for that specific job. Use a design app to lay type over the generated image instead.

Default personalization is the most impactful V7 feature most users never fully activate. The setup requires rating about 200 image pairs, which takes five to ten minutes. The model then builds an aesthetic profile — it learns whether you prefer photorealistic over stylized, bright and clean over dark and atmospheric — and subtly tunes every subsequent generation toward your taste.

Results are mixed, honestly. One designer reported that personalized V7 consistently produced images closer to their brand aesthetic without extensive prompting; another found the personalization too subtle to notice. The difference seems to depend on how polarized your aesthetic preferences are. If you rate images consistently — reliably preferring a clear style — the model has strong signal to work with. If you rate somewhat randomly, the profile will be weak.

Set it up regardless. Spend the ten minutes rating images from genres you actually care about. Once active, use --p [your personal code] to apply your profile explicitly, or just leave personalization on in settings and let it run in the background.

These are the parameters worth knowing. V7’s handling of some has changed from V6 — specifically, --weird no longer accepts values; attempting it returns an error. Keep chaos values between 5–30 for controlled outputs. Above 30, results get genuinely unpredictable.

Source: Medium cheat sheet, Dec 2025 and DataCamp V7 guide, Apr 2025.

The biggest mistake V6 users make when switching to V7 is prompt length. In V6, longer prompts gave the model more to work with. In V7, the model interprets energy and mood as much as it interprets literal content. A 40-word prompt that builds a coherent feeling often outperforms a 120-word prompt that catalogs every visual element.

That said — direction still matters. Here are the structural patterns that consistently work:

1. Lead with the core subject and its relationship to space. “A figure at the edge of a cliff” gives the model spatial grammar before you add anything else.
2. Add lighting as a secondary clause, not a descriptor. “Late afternoon, rim light from behind” rather than “backlit figure.” Specific light placement produces more consistent cinematic results.
3. Set mood through reference, not adjectives. “Like a 1970s National Geographic photograph” outperforms “dramatic, evocative, atmospheric.” References give the model a concrete aesthetic target.
4. Use --no for exclusions. If you don’t want lens flare, text, or specific elements cluttering the frame, --no text, lens flare is cleaner than trying to avoid them in the positive prompt.
5. Lock composition with seeds during iteration. Once you find a composition you like, note the seed. Future variations from that seed will maintain the rough spatial layout while letting you change style or detail.

One oddity worth flagging: adding “paparazzi” to a portrait prompt creates a candid, unstaged quality — useful for lifestyle photography aesthetics. Strange, but it works reliably. V7’s training data means prompt words carry associative visual weight beyond their literal meaning.

Before generating a single image, do these three things. They take fifteen minutes and will improve every output you produce afterwards.

1. Rate 200 images to build your personalization profile. Go to midjourney.com, find the “Rate Images” section, and work through the pairs. Be decisive — consistent strong preferences build a better profile than hedged ones.
2. Set your default model to V7 in /settings. Also set Style to “Med” as your starting point. Starting at “High” produces uncalibrated outputs with V7’s opinionated aesthetic engine.
3. Switch to the web app for serious work. The Discord interface is fine for quick experiments. The web app gives you access to the image editor, inpainting, and Vary Region in a much more usable UI.

Midjourney currently offers four subscription tiers, all paid (the free trial is suspended indefinitely). Annual billing saves 20% across all tiers.

The honest recommendation: start with Standard. The unlimited Relax Mode is the real differentiator — once your fast hours run out, you can keep generating at slower speeds without spending more. Basic users hit a hard wall at ~200 images and can’t generate until the month resets. For most people, that wall arrives sooner than expected.

If you’re doing commercial work and your company makes over $1M in gross annual revenue, the Midjourney Terms of Service require the Pro or Mega plan — not just recommend it. Worth knowing before you publish at scale.

Competitor Comparison: Where V7 Wins and Loses

Pricing sourced from official plan pages; competitive characteristics based on hands-on testing and AI Tool Analysis review (Nov 2025).

What Doesn’t Work (Honestly)

Every guide on this topic skips the failure modes. Here are the real limitations you’ll hit:

Complex text inside images. Despite the claimed 15% improvement, V7’s in-image text accuracy for complex sentences sits around 30%. For any output where readable text matters — product mockups with labels, infographics, signage — generate the image without text and add type in a design tool. Don’t fight the model on this.

Personalization drift. The personalization profile shapes outputs in ways that aren’t always predictable. If you find that V7 keeps pushing outputs toward a particular aesthetic you’re trying to avoid, temporarily disable personalization with --no p or reset your profile. The profile is built from your rating history and will reflect whatever patterns you’ve reinforced.

The learning curve is real. Mastery of V7 takes 20–30 hours of experimentation, not twenty minutes. The first 50 images you generate will likely feel generic. That’s normal. You’re learning to speak a visual language, and that takes repetition. Build a prompt library of what worked and what didn’t; note the seed numbers for compositions you want to revisit.

Where This Is Going

In June 2025, alongside making V7 default, Midjourney launched video generation — producing 5–21 second clips from static images or text prompts. Video consumes significantly more GPU time than stills — roughly 8× the cost per generation according to independent estimates — so it’s not practical at Basic or Standard plan levels if you need high volume. But the capability exists and will mature.

The competitive landscape shifted in 2025. ChatGPT’s image generator (DALL-E 3) caught up significantly on prompt adherence and text rendering. Midjourney’s moat is now specifically aesthetic quality and mood — not general image generation. That’s a narrower moat. It’s also a defensible one, because the artistic quality ceiling that V7 can hit is still above any direct competitor for purely visual work.

Expect V8 to push further on consistency across sequences — the remaining pain point for anyone trying to build visual narratives or maintain character identity at scale. Whether that arrives in late 2025 or 2026 depends on how quickly Omni Reference feedback shapes the training priorities.

The Bottom Line

V7 is a genuinely better model than V6 — sharper anatomy, stronger mood coherence, smarter prompt interpretation. The new features (Draft Mode, Omni Reference, personalization) are worth learning. They require specific setup and adjusted habits, not just better prompts.

The one thing this guide can’t substitute for: time generating images. Read the parameters, understand the features, then go spend two hours making things that don’t look how you expected. That’s where the actual learning happens.