THE RESEARCH / OUR INTERPRETATION
Midjourney’s documentation explains that image references guide a new creation rather than reproduce the source exactly. Our practical implication: an image prompt is not a geometry lock.
Write a change brief before a visual prompt
Consider an illustrative reception-area study: the client wants a warmer material palette, while the approved counter, door opening and accessible route must remain unchanged. Save a named source view and list the permitted changes separately from the fixed elements. “Warmer oak and a quieter stone finish” describes a direction. It does not authorise a deeper counter, a narrower doorway or an invented lighting recess.
- 01Source view
- 02Permitted changes
- 03Fidelity check
- 04Client comparison
Illustrative workflow. The reference photography on this page is not a before-and-after result.
A reception study with an explicit design contract
For a concrete workshop exercise, take a reception with an approved plan, one eye-level model view and a client request to replace a cool grey palette with oak, mineral plaster and a darker counter finish. The design contract fixes the counter footprint, finished floor level, door openings, ceiling line and camera. The permitted variables are finish, texture scale and broad colour relationships. Three source files have different jobs: the plan establishes spatial constraints, the model view establishes visible geometry and a material reference establishes visual intent. A photograph of another interior may help explain atmosphere, but it cannot overrule the project geometry. Keeping these roles separate makes the prompt shorter and the review more objective. It also reveals when the brief is incomplete before image generation starts.
Choose the method for the decision
Use generative studies when discussing atmosphere and broad material relationships. For an option that depends on exact geometry, return to the approved 3D model and produce a render with controlled cameras and material assignments. Local image edits can be useful between these steps, but review their edges and adjacent objects. More prompting is not always the answer to a fidelity problem. Sometimes the correct move is to change the production method.
Prompt structure is a small production specification
A reusable prompt is better organised as a specification than as a chain of adjectives. It identifies the source revision, the intended use of the image, the allowed changes, the invariants and the expected deliverable. For this exercise: “Develop a material-direction study from view R03-CAM02. Explore natural oak joinery, mineral plaster and a dark stone counter finish. Preserve the visible openings, counter silhouette and camera. Output one comparison view; do not add furniture.” This expresses an instruction, not a guarantee that a model obeys it. A separate review record should identify the model or tool, the settings exposed by that tool, reference files and selected output. Prompt history without these inputs is insufficient to reconstruct why an image changed. Reproducibility may still be limited by changes in the service.
| Element | Allowed change | Review evidence |
|---|---|---|
| Counter | Finish only | Silhouette + plan reference |
| Door opening | None | Source-view overlay |
| Timber joinery | Tone + grain appearance | Panel direction + scale |
| Camera | None | Viewpoint + crop |
A visual review does not verify dimensions, product availability or material performance.
Review against an invariant checklist
Compare the source and candidate side by side. Check openings, counter proportions, circulation, junctions and camera position before judging the palette. Record each discrepancy as a correction, an accepted exploration or a reason to reject the image. “Looks realistic” is not a review criterion. A convincing stone texture can distract from a missing support or a changed ceiling height. Dimensions must come from project information, not pixels.
Use an overlay to expose design drift
The first review pass concerns geometry. With an unchanged camera and crop, a semi-transparent overlay can reveal a moved door jamb, altered ceiling line or swollen counter. It is a visual discrepancy check, not a dimensional measurement: differences in lighting and material edges can produce misleading signals. The second pass concerns material logic. Does grain direction follow the joinery? Does a stone pattern imply slab sizes or joints that the design has not resolved? Has a reflective finish created a fictitious opening? The third pass concerns presentation. Is the intended comparison legible, and are the alternatives genuinely different enough to support a decision? Separating these passes prevents a strong atmosphere from masking a geometry error. If the camera itself drifts, the candidate usually needs replacement rather than increasingly elaborate overlay correction.
Present the decision, not the whole image batch
Select a small comparison with the same viewpoint and a short explanation of what changes. Label exploratory imagery and separate material intent from a specified product: finish, availability and performance still need verification. Keep the source revision and review notes with the selected views. In a workshop, the useful takeaway is this repeatable comparison method, not a promise that a particular model will preserve every detail.
Know when to move back into the model
The handover to controlled modelling should occur when the question becomes specific: a stone joint must align with a drawer, a veneered panel needs a defined grain direction, or a finish must be evaluated under a particular lighting arrangement. At that point, the selected image becomes a reference attached to a decision, while geometry, materials and camera settings return to an editable model. The workshop deliverable can therefore be modest but useful: one reviewed comparison board, an invariant checklist, a prompt specification and a record of what still requires modelling or supplier confirmation. There is no need to present dozens of outputs as progress. The relevant measure is the number of usable design decisions reached with acceptable correction effort, not the number of images generated.