A Reusable AI Image Workflow, From Brief to Final File

This guide gives you a repeatable image pipeline: what each step produces, how to judge whether it is done, and why most rework actually happens in the first two steps.

The six-step pipeline

Break a job into six steps: brief breakdown, prompts and references, batch drafts, selection, local repair, upscale and size normalisation, final layout. Layout is the seventh step, but it is what the first six serve.

Every step should have a concrete output. Brief breakdown produces purpose, ratio, must-have elements and must-avoid elements. The prompt stage produces one reusable template plus one or two references. Drafting produces four to eight composition options. Selection produces one master frame. Repair produces a clean subject. Sizing produces files that match platform specs.

Most people skip step one. Without a purpose and a ratio, every later judgement lacks a criterion, so images get chosen on looks alone, and then fail to fit the layout, forcing a regeneration.

Lock the aspect ratio first

Ratio follows use, not taste. A WeChat article header runs around 2.35:1, Xiaohongshu posts are 3:4, slides are 16:9, marketplace main images are 1:1, and single social images are often 4:5 or 1:1.

Once the ratio is fixed, plan the safe area inside the prompt: subject left or right of centre, empty space above or below, nothing important near an edge that may be cropped. Feeds crop again, so the same cover may only show its centre in a list view. Keep the subject comfortably inside.

Order matters. Generate a square and crop later and you will cut the subject. Generate narrow and extend later and the composition changes, which is a rebuild. Thirty seconds spent on ratio saves a whole generation round.

The master-image method

A series is not made by writing every prompt carefully. It is made by writing one carefully and then changing only the variables.

The master is a frame whose composition, light and style all work, plus a complete record of its prompt and settings. Split it into fixed and variable parts:

• Fixed: style words, lighting words, lens words, ratio, model version, style reference.
• Variable: subject, action, scene detail, colour.

For a set of six, generate the master first. Once approved, copy the prompt, change only the variables, and run the batch. That is how six images end up looking like one set. Rewriting each prompt guarantees drift, because your wording shifts every time.

Three repair tools

When a frame is not right, diagnose the problem before rerolling.

1. Local redraw: hands, expression, object position or material are wrong. Repaint only that region and keep the composition you already chose.
2. Outpainting: wrong ratio, need space for a headline, vertical must become horizontal. Extend outward and let the model complete the edges.
3. Erasing: clutter, passers-by, leftover text, blemishes. Mask and remove; the background fills in.

The rule is simple: repair locally whenever possible. A full regeneration throws away the composition you selected, and composition is the hardest part to reproduce.

Three checks before delivery

A repaired image is not a deliverable. Run three checks:

• Resolution: article covers should be at least 1080 pixels wide, and print work should be calculated backwards from 300dpi. Do not stretch a screenshot.
• Ratio: exactly matches the final layout. A small mismatch forces a second crop, and that is where white edges and misalignment come from.
• Text space: the subject must not sit under the headline, and at least twenty percent of the frame should stay clear, with room for either white or black type.

Two more worth adding: scan the edges for AI artifacts such as extra fingers, merged objects or garbled text, and check saturation on a phone. Colour that looks right on a monitor often looks loud on mobile.

Sizes and safe areas by channel

• WeChat header: very wide, subject centred, key information away from the right-hand title zone. Also prepare a square version for list views.
• Xiaohongshu: 3:4 vertical. The first cover decides the click, so the subject should fill more than half the frame.
• Slides: 16:9, with a title zone at the top or left, and no image over the footer logo.
• E-commerce detail pages: 1:1 or 3:4, subject centred and slightly high, with selling points below.
• Short-video covers: generate 16:9 and 9:16 separately. In vertical, keep the subject in the upper half and reserve the bottom third for the title and subtitles.

Produce multiple ratios rather than hard-cropping one file. Name files as date_purpose_ratio_v1 so the final version is obvious once versions pile up.

Treat prompts as team assets

Solo work runs on feel; team work runs on records. For each project, keep one folder with four things: the prompt, the parameters, the references, and the generation date plus model version. Miss any one and nobody can reproduce your result.

Handing a new colleague the master prompt is more useful than handing over ten finished images. Drift comes from rewriting each time; the fix is a fixed template plus a variable table, maintained like part of the design system.

One rhythm to finish with: spend half an hour a month reviewing which prompts were stable and which kept failing, and write the conclusions back into the template. A workflow is not built once; it is worn into shape by projects.

If you work mainly in Midjourney, see the Midjourney prompt guide. If you use domestic tools, see the Jimeng image guide. For more walkthroughs, browse the tutorial list.