AI Video Workflows10 min read

Luma "Modify vs Create" Explained for Creators: A Practical Decision Tree (and How to Mirror It in Veo3Gen)

A practical decision tree for Luma Modify vs Create, plus a worked prompt example and a Veo3Gen workflow to reduce drift and ship faster.

On this page

TL;DR

Use Modify when the composition is already right and you want targeted changes (lighting, color, small props). Use Create when you need a new composition—even if you’re “matching the style” of a reference.

Mirror the same logic in Veo3Gen:

  • If you’re still designing the shot → regenerate (text-to-video).
  • If you’re polishing a shot that’s already right → iterate from an existing frame (image-to-video) and, when needed, pin continuity with first-and-last-frame control on Veo 3.1.

Key takeaways

What Luma “Modify” vs “Create” actually means

Luma’s prompting guide calls this “the first decision”: create something new or change something that already exists. (https://docs.agents.lumalabs.ai/guides/images/prompt-guide)

Create = new composition

Luma describes Create (type: image) as generating a brand-new composition where none of the input pixels are preserved. (https://docs.agents.lumalabs.ai/guides/images/prompt-guide)

Creator translation: you’re commissioning a new shot.

Modify = edit this specific image

Luma describes Modify (type: image_edit) as editing a specific input image where composition is preserved unless you ask for a change. (https://docs.agents.lumalabs.ai/guides/images/prompt-guide)

Creator translation: you’re doing finishing work on an approved frame.

The practical decision tree (60 seconds)

Use this before you iterate.

  1. Is the current framing/composition already the one you want to ship?
  • Yes → go to #2
  • No → Create
  1. Are your changes primarily “surface” changes (lighting, time-of-day, color, texture, mood)?
  • Yes → Modify
  • No → go to #3
  1. Does the change require new staging (big pose/action change, different camera distance/angle, new layout)?
  • Yes → Create
  • No → Modify

Rule-of-thumb examples straight from Luma:

Prompting differences that actually matter

Write like a collaborator, not a keyword blender

Luma’s best practices recommend natural language, describing what you want in detailed language. (https://lumalabs.ai/learning-hub/best-practices)

Also: Luma’s prompting guide says negative prompts aren’t supported, so don’t spend effort on “no blur / no artifacts.” Put that effort into a clearer target description. (https://docs.agents.lumalabs.ai/guides/images/prompt-guide)

Use the “Fast Start” fields as a checklist

Luma’s guide includes a “Fast Start” template with fields for subject, style, lighting, camera/composition, environment/background, mood, details. (https://docs.agents.lumalabs.ai/guides/images/prompt-guide)

How you fill those fields changes by mode:

  • Create: you’re specifying the whole shot → fill every field that matters.
  • Modify: you’re specifying the delta → lead with invariants (what must stay), then describe the change.

Worked example (with before/after you can copy)

Scenario: you already have a strong product hero frame (sneaker on a pedestal). You want the same composition but a different look: “rainy neon night.”

Step 1: Decide the mode

  • Composition is already right.
  • Changes are lighting/mood/surface reflections.

Decision: Modify (composition preserved unless you request a change). (https://docs.agents.lumalabs.ai/guides/images/prompt-guide)

Step 2: Compare prompts (why one drifts)

Goal Prompt What goes wrong/right
“Rainy neon night” but same shot Bad (Create-ish): “Create a cinematic neon rainy night product shot of a sneaker, cyberpunk lighting, shallow depth of field.” You implicitly asked for a new composition; the model can restage camera angle, pedestal, and product placement.
“Rainy neon night” and keep the shot Good (Modify delta): “Modify this image: keep the same camera angle, framing, pedestal shape, and sneaker design. Change lighting to rainy neon night with wet reflections on the pedestal, blue–magenta rim light, and light mist in the background. Keep the logo and laces sharp.” You protect invariants first, then describe the intended change in natural language (no negatives).

Why this aligns with Luma’s guidance:

The iteration loop: Start → Direct → Refine → Finish

Luma’s prompting guide lays out Start → Direct → Refine → Finish. (https://docs.agents.lumalabs.ai/guides/images/prompt-guide)

Use it as a strict four-pass limit so you don’t spiral:

  1. Start: one sentence: subject + what you’re making.
  2. Direct: lock camera/composition + 2–3 non-negotiables.
  3. Refine: fix one major failure only (pose, product readability, mood).
  4. Finish: add small details after stability (atmosphere, micro texture).

Reference images: when they help vs when they trap you

They help when you need identity or style continuity

Luma lists advanced tools including Character Reference and Visual Reference. (https://lumalabs.ai/learning-hub/best-practices)

Luma also notes Dream Machine retains context within a board and “remembers” earlier generations to build upon them. (https://lumalabs.ai/learning-hub/best-practices)

They sabotage when you’re trying to change composition

Early warning signs:

  • You keep getting the reference’s camera angle even when you didn’t ask for it.
  • Background elements “stick” across rerolls.
  • You’re trying to restage, but the model keeps snapping back.

Fix: stop after attempt #2.

  • If the composition is wrong → switch to Create and rewrite camera/composition clearly.
  • If the composition is right but details drift → switch to Modify and list invariants first.

Common failure modes (fast debugging)

1) “It changed everything”

Likely cause: you used Create when you meant Modify. Luma is explicit that Create preserves none of the input pixels. (https://docs.agents.lumalabs.ai/guides/images/prompt-guide)

2) “My prompt feels ignored”

Rewrite in intent-first natural language. Luma says Uni-1 doesn’t pattern-match keywords; it reasons through intent before generating a single pixel. (https://docs.agents.lumalabs.ai/guides/images/prompt-guide)

3) “Pose/action won’t change”

If the action requires restaging, you’re fighting the premise of Modify. Switch to Create for a new composition.

Mirror the decision tree in Veo3Gen (design vs polish)

You can apply the same logic to video so your shots don’t drift across iterations.

Choose your generation approach

  • Design a new shot (Create logic): use text-to-video in Veo3Gen and write an explicit brief (subject + environment + camera/composition + motion + audio intent).
  • Polish a shot that’s already right (Modify logic): use image-to-video from your best frame, and when continuity matters, use first-and-last-frame control on Veo 3.1 to constrain how the shot starts and ends.

Why this is especially practical in Veo3Gen

Veo3Gen generations include native, synchronized audio (dialogue, SFX, music) in a single pass, so you can judge picture + sound together instead of bolting audio on later.

Mid-article CTA (benefit-led): If you’re tired of approving visuals and then reworking the cut because the audio vibe doesn’t match, try a Veo3Gen run where you include a one-sentence audio intent in the prompt—then iterate from the best frame via image-to-video.

Audio intent line you can reuse

Add one sentence that specifies what you want:

  • “Soft room tone and subtle neon buzz; no dialogue.”
  • “Clean product whoosh SFX and light music bed; no voiceover.”

This matches the broader prompting guidance: describe the target outcome in natural language. (https://lumalabs.ai/learning-hub/best-practices)

When to pick Veo 3.1 Fast vs Quality vs Lite in Veo3Gen

Veo3Gen offers three modes:

  • Veo 3.1 Fast (quick, great default)
  • Veo 3.1 Quality (max fidelity)
  • Veo 3.1 Lite (cheapest, preview)

Use the mode choice as part of the same workflow:

  • Early exploration (many rerolls) → Lite/fast iteration.
  • Locking a near-final shot → Fast or Quality, depending on how much fidelity you need.

Checklist

FAQ

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

If you want the same composition with targeted changes, use Modify; if you need a new composition, use Create. Luma describes Create as preserving none of the input pixels, while Modify preserves composition unless you ask to change it. (https://docs.agents.lumalabs.ai/guides/images/prompt-guide)

Can I use negative prompts to stop artifacts or keep things from changing?

Luma’s prompting guide says negative prompts aren’t supported; describe what you want instead. (https://docs.agents.lumalabs.ai/guides/images/prompt-guide)

What’s the fastest way to change time-of-day without restaging the whole scene?

Use Modify and explicitly preserve camera angle/framing while describing the new lighting. Luma’s guide gives “Make this photo look like nighttime” as a Modify rule-of-thumb. (https://docs.agents.lumalabs.ai/guides/images/prompt-guide)

How do I keep the same character across iterations in Luma?

Use Character Reference: upload your image and type @character before your prompt. (https://lumalabs.ai/learning-hub/best-practices)

How do I use a reference image for style without copying the scene layout?

Choose Create for a new composition and use style guidance; Luma’s rule-of-thumb is “Create a new scene in the style of this photo” → Create. Visual Reference is described as uploading an image and using @style before the prompt. (https://docs.agents.lumalabs.ai/guides/images/prompt-guide) (https://lumalabs.ai/learning-hub/best-practices)

How do I reduce drift when doing the same concept in Veo3Gen?

Use the same split: regenerate (text-to-video) while you’re still designing the shot; iterate from an existing frame (image-to-video) when you’re polishing. For tight continuity, use first-and-last-frame control on Veo 3.1.

Closing CTA

If your current workflow is “reroll until it randomly behaves,” the fix is almost always earlier: pick the right mode (Create vs Modify), then write the prompt like a brief with invariants and intent.

When you’re ready to apply that same discipline to video, Veo3Gen is an affordable way to access Google’s Veo 3.1 video models without Google’s enterprise pricing—and it generates video plus synchronized audio in a single pass. Use the free credits to test your first decision-tree run, then scale with pay-as-you-go credits or an optional monthly plan (purchased credits don’t expire).

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.

Sources

Limited Time Offer

Try Veo 3 & Veo 3 API for Free

Experience cinematic AI video generation at the industry's lowest price point. No credit card required to start.

Get Started FreeView Cheapest Pricing
← Back to all posts