Tram cong nghe AI
For photorealistic results in Midjourney V7, write a short prompt that names the subject, the medium (“photo”), the environment, the lighting and the framing. Then add --v 7 --raw plus an aspect ratio (--ar), and keep --stylize low so the image follows your words. Use --no to remove unwanted elements and --oref to keep the same person or object across images. Since July 24, 2026 the default model is V8.2, so you must type --v 7 to use V7.
This guide gives you copy-paste prompt templates for photorealistic portraits, product shots and architecture in Midjourney V7. It also lists the parameters that matter for realism, with the values and limits taken from the official Midjourney documentation.
Some important notes before you start:
- Midjourney runs only on its website and in its Discord bot. There is no official API or command-line tool, and the Terms of Service forbid automated tools and third-party apps. Any “Midjourney API” script you find online is unofficial and can get your account blocked.
- The prompts below are starting templates. Results change with every run, so this guide does not show any “benchmark” or sample outputs.
To find out whether Midjourney is worth paying for, read our Midjourney review.
Which Midjourney version should you use for photorealism?
According to the Version page:
- V6.1: released July 30, 2024; default until June 16, 2025.
- V7: released April 3, 2025; default from June 17, 2025 to June 9, 2026.
- V8.1: released April 14, 2026; default from June 10 to July 23, 2026.
- V8.2: default since July 24, 2026.
You can still pick V7 by adding --v 7 (or --version 7) to the end of a prompt, or by choosing it in your creation settings. Why pick V7 over the newer default? Because some features exist only in V7. The most important one is Omni Reference (--oref), which works only with V7. In V8.x it has been replaced by the Edit Model.
Anatomy of a photorealistic prompt
The Prompt Basics page says that short, clear prompts work best, and that long lists of instructions can confuse the model. It suggests you think about these parts:
- Subject: who or what is in the image.
- Medium: for realism, say “photo” or “photograph” directly.
- Environment: where the scene takes place.
- Lighting: soft, ambient, overcast, neon, studio lights.
- Color and mood.
- Composition: portrait, headshot, close-up, bird’s-eye view.
Parameters always go at the end of the prompt, after the description. A simple pattern that follows this structure:
[medium] of [subject], [environment], [lighting], [color/mood], [composition] --v 7 --raw --ar [w:h] --s [0-1000]Parameters that matter for realism (V7)
| Parameter | What it does | Values (from the docs) |
|---|---|---|
--raw | Turns off Midjourney’s automatic “creative touch”, so your prompt text has more influence. This is the key switch for realistic photos. | Flag only; works in V5.1 and later |
--stylize / --s | Low values follow your prompt closely. High values add more artistic interpretation. | 0–1000, default 100 |
--ar / --aspect | Sets the aspect ratio. Without it, every image is a 1:1 square. | e.g. 4:5, 3:2, 16:9 |
--chaos / --c | Adds variety between the 4 images in a grid. High values may drift away from your prompt. | 0–100 |
--no | Lists things you do not want in the image. | Separate items with commas |
--q | Uses more GPU time on the first grid of images. It does not affect variations or upscales. | V7: 1 (default), 2, 4; --q 3 is changed to 4 |
--weird / --w | Makes the image quirky and unusual, which usually works against realism. | 0–3000, default 0 |
--draft | Runs a V7 job at half the GPU cost. Use it to test ideas quickly. | Flag only |
--oref + --ow | Omni Reference: places a person, object or creature from a reference image into a new image. | V7 only; --ow 1–1000, default 100 |
Sources: Raw, Stylize, Aspect Ratio, Chaos, No, Quality, Weird, Draft Mode, Omni Reference.
A trap with --no
Midjourney reads each word after --no as something to remove. According to the No page, --no modern clothing is read as “no modern” and “no clothing”, and that can trigger a moderation warning. Use single, clear nouns, such as --no text, watermark, people.
Copy-paste photorealistic prompt templates
Replace the parts in brackets. Run each template with --draft first to check the composition cheaply, then run it again without --draft (or with --q 2) for the final image.
1. Editorial portrait
Editorial portrait photo of [a woman in her 30s with short curly hair], wearing [a dark wool coat], in [a quiet café by the window], soft window light from the left, muted warm colors, calm mood, headshot --v 7 --raw --ar 4:5 --s 50Why these choices: --raw and a low --s keep the result close to the prompt, and 4:5 is a portrait frame. If the 4 images look too much alike, add --c 10.
2. Product photo
Product photo of [a handmade ceramic coffee mug] on [a walnut table], soft side light from a window, subtle shadow under the mug, shallow depth of field, close-up --v 7 --raw --ar 4:5 --s 75 --no text, logo, peopleUsing --no text, logo helps prevent random lettering on packaging. Change the aspect ratio to 1:1 for marketplace listings, or to 16:9 for website banners.
3. Architectural exterior
Architectural photo of [a single-story community library with a weathered steel roof and full-height glass walls], [facing a public garden with mature trees], overcast daylight, muted colors of rust, raw concrete and green, wide shot at eye level --v 7 --raw --ar 16:9 --s 100 --no people, cars4. Street photography scene
Street photo of [an elderly fruit vendor at a night market in Hanoi], neon and lantern light, wet pavement reflections, vibrant colors, candid mood, medium shot --v 7 --raw --ar 3:2 --s 1005. Food photo
Food photo of [a bowl of beef pho with fresh herbs], on [a rustic wooden table], soft morning light, steam rising, warm natural colors, top-down view --v 7 --raw --ar 1:1 --s 75 --no textKeeping the same person or product across images: Omni Reference
Many old guides mention --cref (Character Reference). That parameter works only with V6. It was replaced by Omni Reference in V7. Here is how to use Omni Reference:
- On the website, drag your reference image into the prompt bar and place it in the Omni Reference slot. In Discord, add
--oreffollowed by the image URL. - Write a prompt that describes the new scene.
- Adjust
--ow(1–1000, default 100) to control how closely the result follows the reference. The docs note that a high--stylizeor--expcompetes with the reference, so you may need a higher--owwhen you use them.
Photo of the woman walking along a beach at sunset, golden backlight, candid, full-body shot --v 7 --raw --ar 4:5 --oref [image URL] --ow 150Limits from the docs:
- Omni Reference costs 2x the GPU time of a normal V7 image.
- It does not work with Fast Mode, Draft Mode and Conversational Mode, or
--q 4. - It works only with V7. In V8.x, use the Edit Model, which accepts up to 4 reference images.
What about video?
Midjourney video is not part of V7 prompting. It is a separate Video feature (first released as the V1 Video Model), and it works like this:
- You create an image, then click Animate. You can also upload your own image as a start frame.
- Choose automatic or manual motion (you write a motion prompt), and low or high motion.
- Each clip is about 5 seconds long. You can extend it by about 4 seconds at a time, up to four times.
To get a realistic-looking clip, start from a realistic still image made with the templates above.
Common reasons photos look fake (and what to change)
- Glossy “AI art” look: add
--rawand lower--s. - The model ignores some details: the prompt is too long. Cut it down to the subject, setting, light and framing.
- Flat light: name the light source and its direction, such as “soft window light from the left” or “overcast daylight”. The docs list lighting as one of the main prompt elements.
- Wrong framing: set
--ar, because the default is a 1:1 square. - Unwanted text or people: use
--nowith single nouns. - All 4 results look the same: add a small
--cvalue (5–20).
FAQ
Is V7 still available now that V8.2 is the default?
Yes. Add --v 7 to the prompt, or choose V7 in your settings. The Version page has a table showing which features each model supports.
Is --style raw the same as --raw?
The current docs use --raw. Older guides write --style raw, which comes from earlier versions of the docs. Follow the current docs.
Can I generate Midjourney images from a script?
No. Midjourney has no public API, and its Terms of Service forbid automated access. If you need an image tool you can script, look at a local option such as Fooocus.
