Midjourney v7 Complete Guide: New Features and Best Prompts
Everything you need to know about Midjourney v7 including new features, prompt techniques, and parameter settings.
What Actually Changed in v7
Photorealism took the biggest jump. Skin texture, fabric, and natural lighting are markedly better than v6, and hands and faces fail far less often — commonly reported as the single most noticeable upgrade — though fingers in complex grips can still go wrong.
Text rendering improved from nearly useless to usable for short strings. A poster headline of one to four words often comes out clean; a full sentence usually does not. Quote the exact wording in your prompt, for example: a minimalist movie poster titled "NORTH".
Personalization moved from an optional extra to a core feature. Before your first v7 generation, Midjourney asks you to rate a couple hundred image pairs and builds a taste profile that is on by default (toggled with --p). Two people typing the identical prompt now get different results — great for taste, bad for reproducibility. Turn it off when you are following a tutorial or sharing prompts with a team.
Draft mode is the workflow change. --draft renders at reduced quality, commonly reported as roughly ten times faster and around half the GPU cost. It exists to make exploration cheap: iterate in draft, then enhance only the winners to full quality.
Natural Language Beats Keyword Stuffing
v7's language understanding is much deeper than v6's, so full sentences that describe how elements relate to each other outperform comma-separated tag lists.
- BEFORE: "cyberpunk city, neon, rain, night, ultra detailed, cinematic, 8k, trending" - AFTER: "A rain-soaked street market at night in a dense cyberpunk city, vendors sheltering under glowing neon canopies, reflections pooling on wet asphalt, steam rising from food stalls, cinematic wide shot"
The AFTER prompt tells the model how things relate — vendors under canopies, reflections on asphalt, steam from stalls — relationships v7 can actually honor. General prompt anatomy (subject, style, lighting layers) applies here too and is covered in our platform-agnostic image guide; this article stays Midjourney-specific.
Aim for roughly 20 to 60 words. Below ten words v7 improvises heavily; past a hundred, later details get diluted.
Parameters With Practical Values
- --ar (aspect ratio): set it first, because composition changes with the frame. 16:9 for scenes and thumbnails, 2:3 for portraits and posters, 1:1 remains the default.
- --s (stylize, 0-1000, default 100): how strongly Midjourney's house aesthetic overrides your wording. Use 0-50 for photorealism and product shots, 100-250 for general work, 400 and up when you want Midjourney to take creative control. Above roughly 750 it increasingly ignores your prompt.
- --c (chaos, 0-100, default 0): variety across the four images in a grid. Keep 0-10 for client work you need to control, 20-40 while exploring a concept, above 50 only when hunting for surprises.
- --q (quality): leave it at 1 for most work. Prefer draft mode over low quality values for exploration, and reserve higher settings for final renders where your plan supports them.
- --sref (style reference): pass an image URL or an sref code, and control its strength with --sw (style weight, 0-1000, default 100). The commonly recommended band is 100-300 — low enough that your prompt keeps control, high enough that the style visibly transfers. Try "--sref random" to discover looks worth saving.
- --cref (character reference): the v6-era tool for keeping a character consistent. v7 reworked this workflow into its omni-reference system, so if you depend on classic --cref behavior, check the current documentation for your model version. Either way, expect a same-ish person rather than a pixel-perfect identity, and keep reference weights moderate so poses can still change.
A practical starting stack for a stylized scene: --ar 16:9 --s 200 --c 10. For a photoreal portrait: --ar 2:3 --s 50 --c 0.
A Draft-to-Final Workflow
- Explore in draft mode: run five to fifteen cheap variations of the concept, editing the prompt conversationally between runs. - Shortlist and enhance: promote the two or three best drafts to full quality. Expect small changes — enhancement re-renders the image rather than merely sharpening it. - Vary with intent: subtle variations fix small issues; strong variations are for when the composition needs a genuinely different idea. - Upscale last: subtle upscale preserves the image faithfully, while creative upscale reinterprets details and can alter faces. For client work, prefer subtle. - Finish outside: a stray artifact or a single wrong letter is usually faster to fix in an image editor than to re-roll.
Where v7 Still Struggles
- Text beyond a few words: paragraphs, dense UI mockups, and small labels still garble. - Counting and spatial logic: "five dancers" or "the knife to the left of the plate" succeed only inconsistently. - Multi-character scenes: getting two specific characters to interact reliably remains hard even with reference features. - Precise local edits: v7 is a generator, not an editor; targeted changes are better done with region editing tools or external software. - Reproducibility: personalization plus ongoing model updates mean the same prompt drifts over time. Record seeds and full parameter strings for anything you may need to recreate.
Common v7 Mistakes
- Carrying over v6 habits: keyword walls and "8k, hyperdetailed" filler now add noise rather than quality.
- Cranking --s past 700 and then wondering why the prompt is being ignored.
- Leaving personalization on while debugging a prompt, so you cannot tell what the model did versus what your profile did.
- Judging concepts on full-quality renders instead of drafts, burning GPU hours on ideas that were never going to work.
- Skipping --ar and then cropping the best image into a weaker composition.
Treat v7 as a collaborator that reads sentences: describe the scene and its relationships, keep parameters modest, iterate cheaply in draft mode, and spend full quality only on the winners.