Locking a style across a set, using the first frame as the contract

Generate thirteen images from thirteen carefully written prompts and you will get thirteen different worlds. Each may be defensible alone. Placed in a grid on a website they look like a stock library, because the eye reads the differences before it reads the quality.

Consistency is a separate problem from quality and it is solved separately. Writing better individual prompts does not converge them; it produces thirteen better-but-still-unrelated pictures.

What works is treating the first frame as a contract that every subsequent frame is measured against — and being willing to reject a technically better image that breaks it.

Do not start at the beginning of the set. Start with the frame that is most representative — the one whose lighting, palette and treatment you want everywhere.

Iterate on it until it is genuinely right, not acceptable. Twenty attempts on frame one is a good investment; it is the only frame whose quality multiplies across the set.

Then lock it. Write down the exact prompt, the seed if your tool exposes one, and the settings. This is now a reference document, not a draft.

Every other frame in the set is generated with frame one attached as a style reference. The prompt changes; the reference does not.

Weight matters. Full strength pulls composition and subject across too, and you get thirteen variations of the same picture. Somewhere below half typically transfers palette, contrast, grain and light without dragging the arrangement along — but test this once for your tool rather than trusting a number.

The prompts for frames two onward should also be structurally identical to frame one's, with only the subject block changed. Rewriting the treatment or lighting language for each frame is the most common way sets drift while the reference is technically doing its job.

Alongside the reference, keep a block of text that appears verbatim in every prompt in the set:

# shared block, pasted unchanged into all thirteen
Single hard key from upper left, one clean shadow.
Bone-grey seamless ground, matte.
35mm, subject at three metres, sharp throughout.
Kodak 250D grain. Lower third empty.
No logos, no readable text.

The discipline is that this text is copied, never retyped. Retyping introduces small variations — 'hard key' becomes 'strong key', 'bone-grey' becomes 'off-white' — and each one moves the output slightly. Across thirteen frames those movements are visible.

Keep it in a file. It is the set's specification, and it is worth more than any individual prompt in the project.

Review candidates in a grid alongside the frames you have already accepted, never full-screen alone. A frame that is beautiful by itself and slightly warmer than the others is a failure; a slightly duller frame that sits correctly is a success.

A practical test: shrink the grid to thumbnail size and look for the one that jumps out. If a frame draws the eye at thumbnail scale for reasons unrelated to its content — brighter, warmer, more contrast — it has broken the contract and needs re-rolling regardless of how good it looks large.

Reject on set coherence first, then on quality. Doing it in the other order means you accumulate good images that do not fit and have to be replaced later.

No matter how disciplined the generation, thirteen frames will have small differences in exposure and colour temperature. Trying to fix that with more re-rolls is expensive and unreliable.

Bring all of them into one timeline or one editing session and grade them as a set — matching exposure and balance across the group rather than perfecting each individually. This is exactly what a colourist does with footage from a real shoot, and for the same reason.

Fifteen minutes here does more for coherence than another fifty generations. Plan for it from the start rather than treating it as a rescue when the set does not hang together.