Midjourney 8: how to write prompts the model actually understands
MidJourney · Updated:
Midjourney V8 Alpha is the new model, available since March 17, 2026 only on alpha.midjourney.com (Discord is not supported). V8 is not an improved V7 — it's a different model with literal interpretation, native 2K via `--hd`, 5x faster generation, and personalization at the center. The headline rule: state lighting explicitly, otherwise the result will be flat.
What changed in V8
V8 is the shift from artistic interpretation to literal prompt execution. V7 filled in what you didn't specify — added pleasing light, warm glow, artistic treatment. V8 gives you exactly what you describe, nothing more. The prompt «woman in a forest» yields a beautiful frame with warm light in V7; in V8 it produces a technically accurate but flat image.
Practical consequence: state lighting, mood, composition, and style explicitly. Specification beats interpretation. This is not a flaw — it's a tool for users who know what they want.
- Literal interpretation — what's described is what you get
- Native 2K (2048x2048) via --hd
- ~5x faster than V7
- Neutral photographic default output
- Reliable --no for negative prompts
V8 prompt structure
Template: [Subject with details] + [Setting/context] + [Specific lighting] + [Camera/lens] + [Style/mood] --ar X:Y --s N --v 8.
Key rules: 1) State lighting explicitly — V8 won't add nice light for you. Instead of «dramatic portrait, studio lighting» write «Portrait lit from a single softbox positioned 45 degrees to the left, creating a strong shadow across the right side of the face». 2) Write coherent descriptions, not tag lists. 3) Add photographic anchors — V8 responds strongly to «85mm f/1.8», «medium format», «overcast natural light». 4) Lock --ar at the start — changing aspect ratio mid-way equals starting over.
Parameters and modes
Base set: `--v 8` (version), `--ar` (aspect ratio), `--s` (0–1000, default 100), `--c` (0–100, chaos), `--no` (negative prompt — works significantly better than V7), `--style raw` (removes the default polished house style), `--p` (personalization), `--sref [URL]` (style reference, 4x cost in alpha), `--exp` (experimental detail).
V8-exclusive: `--hd` — native 2K (2048x2048), every pixel generated from scratch, 4x cost. `--q 4` — boosted coherence, 4x cost. The `--hd + --q 4` combination works only in Fast mode, not Relax. Use them only for finals.
Working `--s` range in V8: 0–50 for product shots, 50–150 for photorealism (recommended), 200–400 for editorial, 500–800 for free interpretation, 1000 for maximum.
In-image text
V8 renders in-image text significantly better than V7 — signs, packaging, short text (1–3 words). The headline rule: wrap text in quotes.
This works: «A vintage neon sign reading "HOTEL PACIFIC" on a rain-soaked street at night». The model reliably renders that exact word. Without quotes, long text gets mangled — extra letters, scrambled characters.
For more complex text (infographics, long taglines) V8 is still weaker than GPT Image 2 and Nano Banana Pro. If the task is a UI mockup or dense infographic, consider switching to a specialized model.
Common mistakes
1. No explicit lighting
The single most common V8 mistake. «Woman in a forest» yields a technically accurate but flat image. V7 added nice light automatically; V8 doesn't. Always state the light source, direction, and character: «soft morning light from the left», «harsh midday sun from above», «warm backlight creating rim lighting».
2. V7-style prompts (comma-separated tags)
«mountain, fog, sunrise, epic» yields a literally accurate mountain, fog, and sunrise with no artistry. V8 expects coherent descriptions: «A snow-capped mountain emerging from fog at sunrise, dramatic warm light on the peaks, wide-angle composition».
3. Quality-word spam
«beautiful, stunning, 8k, masterpiece, best quality» — V8 ignores these entirely. They don't affect output and burn positional weight at the front of the prompt. Replace with concrete anchors: lens, light type, mood.
4. Using --hd + --q 4 for iteration
Full V8 renders with --hd cost 4x, with --q 4 another 4x, and the combo works only in Fast. These are final-render parameters, not for direction-finding. Iterate in Draft mode at standard quality, switch to --hd only for finals.
5. Personalization without ratings
`--p` is useless if you've rated few images in the Midjourney UI. The model builds the profile from your thumbs up/down — without data, personalization does nothing. Before using --p, rate at least 100–200 frames deliberately (not in a row, with thought).
Before / after examples
Example 1
Before
dramatic portrait, beautiful lighting, 8k, masterpiece
After
A middle-aged man with weathered skin and kind eyes, wearing a worn denim jacket, standing in a doorway of an old bookshop, soft afternoon light from a nearby window casting warm shadows across his face, shot on 85mm f/1.8, shallow depth of field, editorial photography --ar 4:5 --s 100 --v 8
The V7-style «dramatic, beautiful, 8k» yields a flat result in V8 because the model doesn't add nice light automatically. Explicit description of the window, light direction, and shadow radically changes quality.
Example 2
Before
neon sign on rainy street
After
A vintage neon sign reading "HOTEL PACIFIC" on a rain-soaked street at night, reflections on wet asphalt, warm amber glow, cinematic photography --ar 16:9 --v 8
In-image text is one of V8's strengths. Exact word in quotes, setting description, light type. V7 would do this frame worse — V8 renders 1–3 words reliably.
Example 3
Before
product photo of perfume bottle
After
Product photograph of a glass perfume bottle on white marble, single overhead softbox light, clean shadow, commercial photography --ar 1:1 --s 50 --style raw --v 8
For product photography in V8 the `--s 50 --style raw` combo is the best baseline. It disables both artistic interpretation (--s) and the default polished treatment (--style raw). Ideal for commercial shots.
