Build an AI image workflow
you can repeat.

The exercise: one scene, three deliberate variations

You will make a small set of image studies and a record of the choices behind them. The goal is controlled variation, not identical output across different models or services.

You need: an image-generation tool, somewhere to save outputs and a simple text file. A browser service such as Adobe Firefly can be an accessible starting point. ComfyUI provides a node-based approach when you want to work with the steps of a generation workflow. Installation requirements and available controls differ.

This guide supplies an original creative exercise. Its prompts are starting points, not claims of tested results on a particular model.

1. Describe the reference before naming the style

Open the Art Style Database and choose an example that interests you. Write down three visible properties: perhaps a restricted palette, a strong diagonal and a surface that resembles cut paper. Do not stop at “beautiful” or “cinematic.” Neither tells you where the light falls.

Use the reference to develop a vocabulary. Uploading or republishing an existing image is a separate decision: check its provenance and usage rights. The database includes different kinds of styles and concepts, so read the entry’s explanation as well as looking at the picture.

For this exercise, use a simple scene: a blue ceramic cup on a wooden table beside a window. A small scene leaves fewer places for an unexpected object to hide.

2. Write a brief with visible decisions

A blue ceramic cup on a pale wooden table beside a tall window. Morning light enters from the left. The cup sits in the lower-right third; the upper-left area remains quiet and uncluttered. Muted blue, warm cream and soft brown. Visible paper-like texture, simplified shapes, gentle shadows. No lettering, logos or additional objects. Wide composition suitable for an article header.

The subject, light, palette, texture and composition each have a job. If you want a different art direction, change the visual properties rather than accumulating a dozen unrelated style names.

Select a wide aspect ratio in the tool if it provides that control. If the final destination has exact dimensions, record them in the brief and plan a final crop. Asking for a banner in prose does not guarantee a particular pixel size.

3. Save a baseline before chasing improvements

Generate an initial result and save both the image and the exact prompt. Record the model, date, dimensions and exposed settings. If the tool provides a seed, record it too, while recognising that a seed alone does not guarantee reproducibility across versions or environments.

Now inspect the cup’s shape, the handle, the direction of the shadow and the requested empty area. At this stage, technical correctness and composition matter more than how flattering the thumbnail looks.

Write one sentence about what works and one about what fails. “The texture is right; the window dominates the cup” is useful. “Make it better” leaves you with a slot machine and a vocabulary problem.

4. Change one visual decision at a time

1. Study A — light: keep the scene and palette; replace morning light with soft overcast light.
2. Study B — surface: return to the baseline; replace paper-like texture with a smooth painted surface.
3. Study C — framing: return to the baseline; bring the cup closer while retaining the quiet area in the upper left.

Where possible, use an editing or reference feature to preserve more of the initial image. Otherwise, treat the outputs as related studies rather than a controlled scientific experiment. A new generation can change several things even when your instruction changes only one.

Keep all three alongside the baseline. Judge them at the size in which readers will actually see them, including a narrow mobile crop. An elegant composition can lose its subject completely when squeezed into a card.

5. Turn the useful decisions into a reusable record

Save the accepted brief, prompt, settings, output and a short revision note together. With a configurable workflow, also save the workflow using the application’s documented export method and record any models or extensions it needs.

For publication, prepare descriptive alt text based on the final image, not the prompt you hoped would work. Add an accurate AI-image caption where appropriate. Keep the original file and export a web-sized derivative that preserves the visible detail you need.

A compact record might read: “Article banner; blue cup at right; quiet upper left; paper texture retained; overcast variant selected; mobile crop checked.” That is enough to help your future self without requiring a museum catalogue for a cup.

When to move to a more configurable tool

Consider a node-based workflow when you repeatedly need the same sequence of operations or want to inspect the intermediate stages. Stay with a simpler interface if it already gives you the control the project requires. More visible machinery is useful only when you have a reason to touch it.

You have finished when you can explain the chosen image and repeat the process with a new subject. The point is not to eliminate surprise. It is to stop losing the good surprises.