Luma "Modify vs Create" (Uni-1) Explained for Marketers: A Decision Tree to Stop Style Drift + How to Mirror It in Veo3Gen

TL;DR

Luma Modify vs Create is not a style choice—it’s a continuity choice.

Use Modify when the output must remain a version of a specific input image. Uni‑1 Modify edits an input image (type: "image_edit") and preserves composition unless you ask to change it (https://docs.agents.lumalabs.ai/guides/images/prompt-guide).
Use Create when you need a new composition. Uni‑1 Create generates a brand‑new composition (type: "image"); references can inspire, but no input pixels are preserved (https://docs.agents.lumalabs.ai/guides/images/prompt-guide).

For marketers, picking the right branch first prevents “style drift” (the model reinventing what you intended to keep constant) and cuts wasted iterations.

Key takeaways

Decide by what must remain true. If identity/composition must match the input, choose Modify; if you want a fresh composition, choose Create (https://docs.agents.lumalabs.ai/guides/images/prompt-guide).
Write prompts like a brief, not a keyword list. Luma recommends natural, detailed language (https://lumalabs.ai/learning-hub/best-practices), and Uni‑1 explicitly says you don’t need keyword stuffing and that it reasons through intent (https://docs.agents.lumalabs.ai/guides/images/prompt-guide).
Negative prompts aren’t supported in Uni‑1. Don’t rely on “avoid/no” lists as your main control; instead, write positive constraints like “preserve label layout” (https://docs.agents.lumalabs.ai/guides/images/prompt-guide).
Use references intentionally. Dream Machine lists advanced tools like Character Reference (@character) and Visual/Style Reference (@style) (https://lumalabs.ai/learning-hub/best-practices). The concept: identity vs look are different levers.
Mirror the decision logic in Veo3Gen using image-to-video (anchored continuity) vs text-to-video (new composition), plus first-and-last-frame control on Veo 3.1 when you need tighter continuity. Veo3Gen provides three modes—Veo 3.1 Fast, Quality, Lite—and generations include native synchronized audio in one pass.

Modify vs Create (Uni‑1) in marketer terms

Uni‑1 frames the first decision as: are you creating something new, or changing something that already exists? (https://docs.agents.lumalabs.ai/guides/images/prompt-guide)

Here’s the marketer translation:

Modify = “keep my layout/identity, change specific attributes.”
- Uses type: "image_edit" (https://docs.agents.lumalabs.ai/guides/images/prompt-guide)
- Preserves composition unless you request a change (https://docs.agents.lumalabs.ai/guides/images/prompt-guide)
Create = “new scene, same vibe.”
- Uses type: "image" (https://docs.agents.lumalabs.ai/guides/images/prompt-guide)
- Can be inspired by references, but none of the input pixels are preserved (https://docs.agents.lumalabs.ai/guides/images/prompt-guide)

If your team keeps saying “the model won’t listen,” it’s often because you’re trying to force Create to behave like Modify (keep exact geometry) or Modify to behave like Create (invent a new composition).

A decision tree you can actually run in a creative review

Answer these in order—no prompting until you can answer #1 in one sentence.

What must remain identical?

If the answer includes composition, product geometry, label layout, the same person, or the same frame → Modify (https://docs.agents.lumalabs.ai/guides/images/prompt-guide)
If the answer is “nothing needs to match exactly; just match the campaign vibe” → Create (https://docs.agents.lumalabs.ai/guides/images/prompt-guide)

What can change freely? (lighting, environment, camera move, props)
What’s your pass/fail test? (one line)

If you want a simple rule: your acceptance test reveals the correct mode. If your pass condition mentions “preserve,” you’re in Modify-land.

The marketer scenario table (with the Luma choice and the Veo3Gen mirror)

Use this as a pre-flight check.

Scenario (marketing task)	What must stay the same?	Luma branch	Why (grounded in Uni‑1)	Veo3Gen mirror (same logic, different UI)
Upgrade a real product packshot to “premium studio”	Bottle + label layout + framing	Modify	Modify edits a specific input and preserves composition unless asked (https://docs.agents.lumalabs.ai/guides/images/prompt-guide)	Image-to-video anchored to the packshot; describe only lighting/background + subtle motion.
Fix one detail: warped logo / wrong on-pack text	Everything except the error	Modify	You’re changing a specific existing image (https://docs.agents.lumalabs.ai/guides/images/prompt-guide)	Re-run image-to-video from same anchor; change only the failing detail in the brief.
New hero scene for the same campaign (fresh concept)	Only brand tone/art direction	Create	Create is a new composition; references can inspire (https://docs.agents.lumalabs.ai/guides/images/prompt-guide)	Text-to-video “shot card” brief; add a reference only if needed.
10 ad variations with different staging/angles	Brand vibe, not layout	Create	You want new compositions	Text-to-video with a reusable template; vary one variable per version.
Same mascot across multiple scenes	Identity consistent, scenes change	Create + Character Reference concept	Dream Machine lists Character Reference (`@character`) as an advanced tool (https://lumalabs.ai/learning-hub/best-practices)	Keep a consistent visual anchor (e.g., same reference image / start frame) and repeat the same identity description.
Same scene, but background swapped	Layout stays, environment changes	Modify	Composition preservation is the point (https://docs.agents.lumalabs.ai/guides/images/prompt-guide)	Image-to-video; explicitly request background replacement while keeping subject placement.

One worked example (before/after) that stops drift

Problem: “Our product keeps changing shape and label layout”

This happens when you prompt like you want a new ad, but you actually need a version of the same packshot.

Before (vague prompt that invites Create behavior)

“Cinematic premium skincare ad, glossy bottle, studio lighting, shallow depth of field.”

What’s missing:

No statement that the label typography/layout must remain identical.
No statement that composition/framing must be preserved.

After (Modify-ready brief block)

Luma best practices emphasize natural language (https://lumalabs.ai/learning-hub/best-practices), and Uni‑1’s guide provides prompt structure elements via “Fast Start” (subject, style, lighting, camera/composition, environment/background, mood, details) (https://docs.agents.lumalabs.ai/guides/images/prompt-guide).

Copy/paste this structure:

Goal (1 sentence)

Turn this exact product image into a premium studio beauty look while keeping the bottle and label identical.

Constraints (3–6 bullets)

Preserve bottle silhouette, cap shape, and label text/layout exactly.
Preserve current framing and product placement.
Only change: lighting, reflections, and background texture.

Shot specifics (Fast Start components)

Soft key light from camera-left; controlled specular highlights.
Clean gradient background; luxury, calm mood.
Subtle push-in; keep composition stable.

Acceptance test (1 line)

If label typography/layout or bottle silhouette changes, the result fails.

Why this works:

It matches Uni‑1 Modify’s intended use: edit a specific input, preserve composition unless requested (https://docs.agents.lumalabs.ai/guides/images/prompt-guide).

Mid-article CTA (Veo3Gen)

If you want to apply the same “anchored vs new composition” logic to video without adding a separate audio step, try the workflow in Veo3Gen: it generates video with native, synchronized audio (dialogue/SFX/music) in one pass, and supports text-to-video and image-to-video.

Prompting: the brief block (what to include, what to stop doing)

What to include

Uni‑1 describes a workflow of Start → Direct → Refine → Finish (https://docs.agents.lumalabs.ai/guides/images/prompt-guide). You can operationalize that as a repeatable prompt format.

Use this every time:

Goal: “What are we making?”
Constraints: 3–6 bullets (“what must remain true”)
Shot specifics: subject, style, lighting, camera/composition, environment/background, mood, details (https://docs.agents.lumalabs.ai/guides/images/prompt-guide)
Acceptance test: one line pass/fail

What to stop doing

Stop keyword stuffing. Uni‑1 says you don’t need prompt-engineering tricks or keyword stuffing and that it reasons through intent (https://docs.agents.lumalabs.ai/guides/images/prompt-guide).
Stop relying on negative prompts in Uni‑1. Negative prompts aren’t supported (https://docs.agents.lumalabs.ai/guides/images/prompt-guide). If you must prevent a change, say what to preserve.

References: identity vs look (why teams keep mixing these up)

Dream Machine best practices list advanced tools including Character Reference (@character) and Visual Reference (@style) (https://lumalabs.ai/learning-hub/best-practices). Whether or not you use those exact handles elsewhere, keep the mental model straight:

Identity consistency = “same character/product/person” across outputs.
Style consistency = “same lighting/grade/material feel” across outputs.

If you need geometry-level fidelity (packaging, exact typography), style references are usually too weak; that’s when you anchor with Modify (or an image-first workflow).

Iteration: how to fix the last 20% without restarting

Treat iteration like a controlled experiment.

Restate what must not change (one line).
Change one variable per iteration (lighting or wardrobe or background).
Promote constraints into bullets (clearer than a long paragraph).
Give feedback like a teammate. Luma recommends treating the agent like a creative teammate and providing context, clear goals, and feedback (https://lumalabs.ai/learning-hub/get-better-ai-results-in-luma-prompting-like-a-creative-teammate).

Optional tactic when requests are vague: the Luma guidance includes a template that asks the agent to pause and request clarification before creating anything (https://lumalabs.ai/learning-hub/get-better-ai-results-in-luma-prompting-like-a-creative-teammate).

How to mirror “Modify vs Create” in Veo3Gen (practically)

Don’t hunt for the same labels—mirror the decision logic.

If you’d pick Modify in Luma → anchor in Veo3Gen

Use when continuity matters.

Use image-to-video to anchor on a specific frame.
Use first-and-last-frame control on Veo 3.1 when you need the motion bracketed by specific start/end states.
Choose a mode based on your phase: Veo 3.1 Lite (cheapest preview), Veo 3.1 Fast (quick, great default), Veo 3.1 Quality (max fidelity).

Veo3Gen supports 720p, 1080p, and 4K (4K on Fast/Quality), with aspect ratios 16:9 and 9:16.

If you’d pick Create in Luma → text-first shot cards in Veo3Gen

Use when you want a new composition.

Shot card mini-template (pasteable)

Objective:
Subject:
Action:
Scene:
Camera:
Lighting/Mood:
Audio (single pass): dialogue + SFX + music vibe
Acceptance test:

That last line is the drift-killer: it forces you to define what “success” means before you generate.

Checklist

Decide: must match input (Modify logic) vs new composition (Create logic) (https://docs.agents.lumalabs.ai/guides/images/prompt-guide).
Write a one-line acceptance test that makes “same vs new” unambiguous.
Separate identity constraints (character/product) from style constraints (lighting/grade).
Use the brief block: goal + 3–6 constraints + shot specifics + acceptance test.
Iterate by changing one variable at a time; don’t restart when you’re close (Start → Direct → Refine → Finish) (https://docs.agents.lumalabs.ai/guides/images/prompt-guide).
If using Uni‑1: remember negative prompts aren’t supported; use positive “preserve/maintain” constraints (https://docs.agents.lumalabs.ai/guides/images/prompt-guide).
In Veo3Gen: choose image-to-video + first/last frame control for anchored continuity; choose text-to-video for new concepts.

FAQ

How do I know if I should use Luma Modify or Create?

Ask: “Do I need a version of this exact input?” If yes, use Modify (type: "image_edit")—it preserves composition unless you request a change. If you need a new composition, use Create (type: "image") (https://docs.agents.lumalabs.ai/guides/images/prompt-guide).

Why does Create keep reinventing my layout?

Because that’s what it’s for: Create makes a brand-new composition and doesn’t preserve input pixels (https://docs.agents.lumalabs.ai/guides/images/prompt-guide). If layout must stay stable, switch to Modify logic and make “preserve composition” explicit.

If Uni‑1 doesn’t support negative prompts, how do I prevent unwanted changes?

Use positive constraints (“preserve label typography/layout,” “maintain framing”) and limit what may change. Uni‑1 states negative prompts aren’t supported (https://docs.agents.lumalabs.ai/guides/images/prompt-guide).

Can I ask for on-screen text in Luma generations?

Yes—Dream Machine best practices say you can ask for text by specifying the desired wording directly in the prompt (https://lumalabs.ai/learning-hub/best-practices).

How do I mirror Modify vs Create thinking in Veo3Gen?

Use image-to-video when you need anchored continuity (Modify logic). Use text-to-video when you want a new composition (Create logic). For tighter continuity, use first-and-last-frame control on Veo 3.1.

Close: turn the decision tree into a system (and ship faster)

The fastest teams don’t “prompt better”—they choose the correct branch first, then run a consistent brief format with a pass/fail line.

If you want to productionize that same logic for video, Veo3Gen lets you work in text-to-video or image-to-video, generate native synchronized audio in a single pass, and scale outputs via a developer API once your templates are stable.

When you’re ready, apply the checklist above to one campaign and run it end-to-end in Veo3Gen—start with a cheap preview mode, lock the brief, then move up to higher fidelity modes for finals.

Start creating with Veo3Gen

Veo3Gen gives you affordable Veo 3.1 video generation with native audio, up to 4K, and credits that never expire — with free credits to start.

Generate your first video now: Get started
Compare plans and pay-as-you-go pricing: See pricing

Luma "Modify vs Create" (Uni-1) Explained for Marketers: A Decision Tree to Stop Style Drift + How to Mirror It in Veo3Gen

Try Veo 3 & Veo 3 API for Free