Midjourney 7: how to write prompts the model actually understands
MidJourney · Updated:
Midjourney V7 is Midjourney's flagship image model released on April 3, 2025. V7 fundamentally changed prompt writing — the model now understands natural English, and the old «keywords with commas» approach actively hurts results. Opening words carry more weight, photorealism is the best in the family, and personalization is enabled by default.
What Midjourney 7 does well
V7 is the shift from tag-soup to descriptions in a cinematographer's voice. The model dramatically improved human anatomy (hands, bodies, poses, spatial relationships), rendering of text on signs and packaging, and comprehension of full sentences.
Key strengths: best photorealism across the Midjourney family, strong response to photography vocabulary (lenses, apertures, film stocks), a new --oref (Omni Reference) parameter for universal content + style transfer, and advanced --sref with --sw weight for visual style transfer. The --cref parameter from V6 is removed in V7.
- Natural language instead of keyword lists
- Dramatically improved anatomy and spatial relationships
- Best photorealism in the family + text rendering
- Personalization (--p) on by default
- Omni Reference (--oref) — new universal reference
Prompt structure
Optimal hierarchy: [Subject] + [Subject details] + [Context/setting] + [Style/mood] + [Camera/lighting] + [Parameters].
The main subject always goes first. V7 processes tokens sequentially — opening words get maximum weight. «Beautiful cinematic photo of a woman» works worse than «An elderly fisherman with a weathered face, standing on a wooden dock at dawn, documentary photography style».
Write in coherent sentences, not commas between tags. The model understands grammar and context: «A man walking through rain-soaked streets, his reflection shimmering in puddles» beats «man, rain, street, wet, reflections» every time.
V7 parameters
Core set: `--ar` (aspect ratio, default 1:1), `--s` or `--stylize` (0–1000, default 100 — strength of artistic interpretation), `--c` or `--chaos` (0–100, variation diversity), `--w` or `--weird` (0–3000, experimentation).
Style and references: `--style raw` removes the default artistic treatment, `--sref [URL]` transfers visual style from an image, `--sw` (0–1000) controls strength, `--oref` is the new-generation Omni Reference, `--iw` (0–3) weights image references, `--p` applies your personalization profile.
Modes: Draft (0.5x cost, fast) for iteration, Fast (1x) for everyday work, Turbo (2x) for urgent finals, Relax (free) for experiments.
Photorealism and lighting
Lighting is the single most powerful quality lever in V7. Specific descriptions of light radically change the result: «golden hour light casting long shadows across weathered stone», «Rembrandt lighting with soft fill from camera left», «bioluminescent glow illuminating the fog» beat «nice lighting» or «dramatic light».
For photorealistic portraits, use photography vocabulary: «shot on 85mm f/1.8, shallow depth of field», «medium format look, high dynamic range», «35mm film grain, Kodak Portra 400 palette». These give the model technical anchors it responds to strongly.
For product shots combine `--s 25–75` with `--style raw` — this minimizes artistic interpretation and produces a literal result. For concept art and illustrations push `--s` to 300–600.
Common mistakes
1. Comma-separated keyword lists
The headline V7 anti-pattern. «mountain, fog, sunrise, epic, cinematic» produces chaotic results because the model expects coherent description. Write in sentences: subject, what they do, where they are, how they're lit.
2. Quality-word spam
«beautiful, stunning, 8k, detailed, masterpiece, best quality» is useless noise in V7. These words burn precious positional weight at the start of the prompt and don't improve results. Replace them with concrete photographic anchors (lens, film stock, light source).
3. Subject not in the opening words
«Beautiful cinematic photo of a woman» — the subject is third, style stole the weight. V7 processes tokens sequentially and gives maximum weight to the opening lines. Move «who or what» into the first sentence.
4. Using --cref in V7
The --cref (Character Reference) parameter is removed in V7. If you copy a prompt from V6 and leave --cref in, it does nothing. Use --oref (Omni Reference) for universal transfer or --sref for style-only.
5. Conflict between --s and --style raw
High `--s 500` plus `--style raw` pull the model in opposite directions: one says «more artistic interpretation», the other «less treatment». For photorealism the combo is `--s 25–100 + --style raw`; for illustration go high `--s` without `raw`.
Before / after examples
Example 1
Before
beautiful cinematic photo of a fisherman, ultra detailed, 8k, masterpiece
After
An elderly fisherman with a weathered face and silver beard, standing on a wooden dock at dawn, documentary photography style, contemplative mood, shot on Leica M11 with natural morning light, soft mist rising from the water --ar 3:2 --s 100 --v 7
The old prompt buries the subject under style and spams quality words. V7 expects the subject first, with specific photographic anchors (Leica M11, natural morning light) replacing «8k, masterpiece».
Example 2
Before
mountain, fog, sunrise, epic, cinematic, dramatic, beautiful
After
Ancient redwood forest in morning mist, shafts of golden light filtering through canopy, ferns covering the forest floor, wide-angle composition, national geographic photography --ar 16:9 --s 150 --v 7
Comma-separated tag lists are a V7 anti-pattern. A coherent description with specific light, composition, and a genre anchor («national geographic photography») reliably outperforms.
Example 3
Before
product photo of a coffee mug
After
Minimalist product shot of a ceramic coffee mug on a light oak table, soft natural window light from the left, clean composition, editorial style --ar 1:1 --s 25 --style raw --v 7
Product photography is the scenario where low `--s` plus `--style raw` is mandatory. Without them the model adds artistic interpretation that breaks commercial use.
