How I AI: Jamey Gannon's Workflow for Consistent Brand Imagery in Midjourney

In this episode of How I AI, creative director Jamey Gannon demonstrates how she builds a repeatable visual system in Midjourney, repairs selected details with Nano Banana, and creates realistic self-portraits for editorial work.

Jamey does not begin with a long photography prompt. She begins with visual references, tests what the model is actually taking from them, and changes one input at a time until the outputs hold together.

Her process moves from a mood board to individual style references, personalization codes, and a client handoff in Figma. A second pass in Flora or another interface using Nano Banana handles targeted edits that Midjourney does not resolve cleanly.

Build a visual system Midjourney can repeat

The companion consistent brand aesthetic workflow lays out the reference-board, SREF, personalization, and handoff sequence in a form a creative team can reuse.

The goal is not to force every image into one template. It is to create a recognizable visual range that can survive new subjects and compositions.

Start with a mood board, not a paragraph

Jamey collects references in Pinterest or Cosmos before she opens Midjourney. For the demonstration, she wanted a pink, cute, internet-influenced look that did not feel overly feminine. The board included deliberately strange combinations such as a grungy unicorn and a fluorescent fruit dog at a computer.

The board gives the model visual information that would be difficult to describe consistently in words. It also gives Jamey a fixed target against which to judge the first generations.

Isolate the references that carry the style

Jamey first tested the mood board with short subjects such as "beautiful female model" and "astronaut." The outputs were usable images, but their contrast, saturation, and overall treatment did not match the board.

A varied Midjourney mood board can average its signals together. Jamey therefore reused selected images from the board as individual Style References, commonly called SREFs, by dragging them into the Midjourney interface.

The SREF test moved closer to the intended style but introduced a strong green cast. Jamey traced that effect to one reference with a vivid green eye, removed it, and generated again. The change showed which reference was steering the palette.

Layer in a personalization profile

Midjourney personalization profiles are created from repeated image preferences. The resulting code can be applied to later generations as another style signal.

Jamey combined selected SREFs with a profile she calls "late 2025 aesthetic." She uses multiple profiles for different looks, so the code is one layer in the system rather than a universal brand setting.

Change one visible input at a time

Once the references are close, Jamey keeps the text prompts direct. She uses familiar editorial references, camera names, and plain descriptions as shortcuts, then judges the image rather than assuming the prompt behaved as intended.

• Publication references: Terms such as "Dazed editorial photo shoot" or "Vogue" can signal a broad editorial treatment without a paragraph of camera language.
• Camera-name shortcuts: Jamey keeps a list of cameras and tries them as aesthetic cues. In the demo she used Sony RX100, while noting that she had likely generated the camera suggestion with ChatGPT. The useful step is testing the visual effect, not treating the model's camera history as authoritative.
• Plain descriptions: She added details such as "New York skyline can be seen in the window behind it" and "on a matte black leather couch." Each instruction changed one visible element.

When an image reference of a woman blowing bubble gum made Midjourney repeat the bubble, Jamey cropped the bubble out of the reference and tried again. Changing the visual input worked better than adding another negative instruction.

Package the recipe, not just the outputs

Jamey delivers the final system in Figma along with example images in context. The handoff records how the look was produced instead of supplying only a fixed asset folder.

The package includes:

1. The selected Style References and Image References.
2. The personalization profiles and relevant Midjourney settings.
3. The final prompts that reliably produced the demonstrated look.

Clients receive both finished examples and the inputs needed to generate more. Jamey described Midjourney sharing as unfinished, so the Figma file serves as the practical record of the system.

Repair the one detail that breaks the image

Use the targeted image repair workflow when the composition works but one object, hand, or product detail does not. The narrow edit protects the choices that already succeeded.

Midjourney can produce a strong composition with one unusable detail, such as a distorted hand or invented computer. Jamey takes those images into Flora, Higgsfield, or another interface that gives her access to Nano Banana for a targeted edit.

In the demonstration, she kept an image she liked and replaced only its outdated-looking computer:

1. Upload the image: Bring the selected Midjourney output into the editing interface.
2. Target the change: Identify the object or region that needs replacement.
3. Describe the replacement: Name the object, placement, visible angle, and elements that should remain unchanged.

replace the computer she's typing on, on a 2026 midnight Black MacBook Pro. keep the position and the size of the computer exactly the same. only the left side and the keyboard is visible

Jamey also adds "don't change anything else." In her example, the edited download increased from roughly 800 by 800 pixels to roughly 4,000 by 4,000 while preserving the surrounding style. She still checks the result visually rather than assuming the named product or geometry is exact.

Build a reusable self-portrait reference set

The realistic AI self-portrait workflow explains how to capture varied identity references and reuse them across editorial compositions.

Jamey uses AI portraits when she needs a specific expression or composition for an article or social image and does not have a matching photograph.

Her process starts with recognizable source images:

1. Create a varied reference set: Take realistic selfies with different angles, expressions, lighting, and visible details. Jamey includes images that show her teeth because that detail helps the output resemble her.
2. Use the selfies as identity references: Load several source photos into an interface such as Flora or Higgsfield.
3. Generate and remix: Jamey can start from a composition she likes, adjust the expression, and replace the subject's face with her own reference set. In the demo, she turned a Midjourney composition into an annoyed self-portrait for an article.

The result is a library of portraits that can be matched to different editorial ideas while retaining the broader visual system.

Creative direction is the operating system

Jamey treats image generation as creative direction. The important decisions are which references belong together, what changed between generations, and whether the output still matches the intended brand.

Mood boards establish the target. SREFs isolate stronger style signals. Personalization codes add a reusable preference layer. Targeted edits repair specific failures without restarting the whole composition.

A practical first test is a small reference board and two or three simple subjects. Keep the prompt stable, change one visual input at a time, and save the combinations that consistently produce the look you want.

Consistency comes from preserving the decision trail. Save the source board, the references that survived testing, the profile code, the short prompt, and the final repair instructions. A teammate should be able to reconstruct why an image belongs to the system instead of merely imitating its surface.