Lock the visual direction before you generate anything

The first three images of a set are always fine. The trouble starts around image eight, when the wood floor has drifted a shade cooler, the light has moved from afternoon to midday, and a chair has appeared that belongs to a different apartment. Nothing is wrong with any single image. The set has stopped being a set.

This is not a model problem. It is a memory problem: each prompt was written from scratch, so each prompt made its own small decisions.

What a direction actually contains

A usable direction is boring and specific. Ours covers, for every project:

• Palette — hex codes with a role attached. #F2EEE7 as the wall base, not “warm off-white”.
• Materials — an allowed list and a forbidden list. Light oak, matte ceramic, brushed steel; no chrome, no gloss plastic, no marble.
• Light — a colour temperature, a shadow behaviour, and the casts that are banned outright. “Soft daylight, 5000K, long soft shadows, no green cast.”
• Camera — height in centimetres, angle per image role, and how much of the frame the product occupies. “Eye level, 90–110 cm, product fills 60% of frame.”
• Who is in it — if there are people or animals, they are named and described once, and they stay the same person or the same animal across the set.
• Props — a whitelist and a count. Two objects on a counter, not “minimal styling”.

Every one of those lines exists because its absence produced a reshoot at some point.

The test

The direction is finished when someone who has not seen the first twenty images could produce the twenty-first and have it belong. If producing image twenty-one requires looking back at image four and guessing, the direction is still in someone’s head.

That test is also the reason the direction is a document rather than a habit. Habits do not survive a second person, a second week, or a client who returns nine months later for three more images.

Feedback changes the prompt, not the direction

Client feedback arrives in rounds, and it arrives at three different heights:

• On the project — “everything is too cold.” This edits the direction itself, and every subsequent image inherits it.
• On the round — “all of these should be square, warmer, with fewer props.” Applies broadly, but not blindly: a hero on a transparent background should ignore “vary the rooms.”
• On one image — “the cat should be sitting upright, not lying down.” Applies to that shot alone, and overrides whatever the original brief said.

Keeping those three apart is what stops round two from quietly undoing round one. A note meant for one image, applied to the whole set, is how a coherent set falls apart in a single afternoon.

A short checklist before the first generation

1. Palette written as hex, with roles.
2. Materials, allowed and forbidden.
3. Light temperature and shadow behaviour.
4. Camera height and frame occupation, per image role.
5. Characters named, if any.
6. Props whitelisted and counted.
7. The set list itself — which image answers which buyer question.

Twenty minutes on that list is worth more than any amount of prompt polishing afterwards. The prompt is a rendering of the direction; you cannot render a decision that was never made.