Midjourney Tutorial: Reverse-engineer prompts with Describe and lock in a style with SREF

This Midjourney tutorial focuses on solving two things: how to reverse-engineer a prompt when you don’t know how to write one, and how to keep a set of images in the same style over the long term. As long as you master Midjourney’s /describe and --sref, you can turn image generation from “pure luck” into a reusable workflow.

How to get started and prepare: first make sure your Midjourney commands work

In Discord, enter any available Midjourney channel or your own private channel, and make sure you can send commands normally and see the bot return results. When using Midjourney for the first time, it’s recommended to generate an image with a simple prompt first to confirm your permissions and that the queue is working, before moving on to the reverse-engineering and style-locking steps. Doing this helps you avoid mistakenly attributing problems to the prompt.

If you mainly use Midjourney in your own server, remember to check whether the channel allows the bot to speak and whether you’re operating in the correct channel type. Midjourney’s command entry is the same everywhere—the key is that the bot can reply and generate a grid of images.

Reverse-engineer with /describe: turn a “reference image” into an editable prompt

When you only have an image you like (photography, illustration, posters—anything), type /describe directly in Discord and upload the image. Midjourney will return multiple English description options, which are prompt prototypes you can reuse directly. You don’t need to copy them verbatim—prioritize keeping these three parts: “subject + materials/lighting + lens/composition.”

In practice, I recommend: first choose the closest description and generate once, then remove unwanted elements with negative terms (for example, get rid of background clutter, text, or a watermark-like feel). Once you’ve edited the reverse-engineered prompt so that it “doesn’t go off track even after generating three times,” that Midjourney prompt is basically ready to go into your常用 library.

Lock in a style with --sref: build your own style reference library

If you want a consistent look and feel across a series of works, the easiest approach is to add style references in Midjourney: append --sref at the end of the prompt and include one or more links to style images. The style images should ideally express only “brushwork, color palette, lighting and shadow, materials,” and not have an overly dominant subject; otherwise, Midjourney may “borrow” the subject as well.

You can also pair --sref with a weight parameter (tweak it slightly if you find the style too strong or too weak) and treat it like a “style knob.” Once you have two or three stable --sref combinations, Midjourney becomes much smoother to use for brand key visuals, series avatars, and fan art with a unified art style.

Common sticking points and fixes: inaccurate reverse-engineering, style drift, inconsistent outputs

If the description from /describe “looks right but generates way off,” it’s usually because the description includes non-essential scene terms; you can delete location/atmosphere words, keep only the subject and visual language, and then run Midjourney again. When style drift happens, first check whether the images used for --sref are too mixed and whether the style itself is consistent; next, consider reducing the style adjectives in the prompt to avoid conflicting with --sref.

If results vary a lot each time with the same prompt, you can fix the random seed (seed) to improve reproducibility, and try to keep the same aspect ratio and the same parameter set. Treat Midjourney as a “controllable workflow” rather than “gacha,” and your time cost will drop noticeably.