Flux Art — AI made simple, unleash your unlimited creativity
Multi-model AI visual creation and production platform · One account and workspace · Images, video, asset management and OpenAPI
Start Creating →
Flux ArtBlogTutorials › Product Photos Look …

Product Photos Look Inconsistent? Use Up to 14 Reference Images

Anonymous community contributor (alias): Wintergreen Sketch Board Published: Category:Tutorials

Direct answer: style consistency isn't about "magic-word prompting" — it comes down to the reference-image mechanism. In Flux Art, Nano Banana 2 supports up to 14 reference images: build your approved, high-CTR photos into a "reference set" and feed it in, so new products follow that baseline. How to split the 14 slots, how to batch-run, and how to sign off — it's all below.

1. Why Prompts Alone Can't Keep Style Consistent

Feed the model the same style brief ("bright, airy Scandinavian look") and it interprets it slightly differently every time — 10 SKUs end up looking like they came from three different teams. A reference image turns "style" from words into an executable visual baseline — and that's the dividing line between a "whole-store visual system" and just "one nice photo."

2. How to Build a Reference Set (Allocating the 14 Slots)

SlotsWhat to IncludePurpose
6–8 imagesStore-wide style baseline (approved photos of hero products)Sets color tone, composition, and negative space
4–6 imagesCategory-representative photosSets the category's shooting angle (flat lay / on-model / 45°)
0–2 imagesThis batch's actual new-product shotsAnchors the subject itself

Two rules: only approved, high-performing photos get into the reference set; newly approved winners get rotated back in, so the set evolves along with your store.

Product Photos Look Inconsistent? Use Up to 14 Reference Images - Flux Art

Figure 1: Where the Nano Banana series, which powers the reference-image mechanism, sits in the model library (see the official site for the current lineup).

3. The Batch Workflow: Three Constants, One Variable

  1. Fix the model (Nano Banana 2), fix the aspect ratio (lock to 1:1 for any-ratio needs, run portrait as a separate batch), and fix the prompt template;
  2. The only variable is the product name and selling points:

```text

Follow the reference set's overall style and composition, swap the subject for: [new product description],

Keep the color tone/lighting/negative-space ratio consistent, don't add new decorative elements

```

```text

Short version: follow the reference set's style and composition, swap subject to [new product],

keep palette/lighting/spacing consistent, no new decorations

```

  1. After generating each SKU, do any uniform touch-ups (background color, price tag) with inpainting — don't regenerate the whole image;
  2. Review the whole batch once it's done running — don't tweak one image at a time as you go. Drift only shows up when you compare across the batch.

Calibrate on the first batch: before running at full volume, use the reference set to generate 3 "control samples" and place them side by side with the baseline images to compare color tone and negative space. If there's a deviation, adjust the reference set's composition (usually by removing one or two off-style images), then run the full batch once calibration passes — the cost of those three samples saves you a full-batch redo.

Product Photos Look Inconsistent? Use Up to 14 Reference Images - Flux Art

Figure 2: The "model" and "aspect ratio" from the three constants are locked in right here on this control row (see the official site for the current interface).

4. The Consistency Sign-Off Checklist

CheckMethod
Color consistencyMix old and new images in a 3x3 grid — a quick visual scan shouldn't reveal "two different teams"
Composition/negative spaceSubject proportion and horizon-line position line up across every image
Text hierarchyHeadline size and placement stay consistent across the whole series
Subject fidelityColor and material match the actual product, avoiding after-sales disputes

5. Find Your Fit: How Should Your Store Build a Reference Set

Your ScenarioBiggest Pain PointHow to Do It in Flux ArtRecommended Main Model
Apparel product seriesInconsistent shooting stylePut on-model/flat-lay baseline photos into the reference setNano Banana 2
Multi-SKU general merchandiseDrift across batchesRun batches with three constants, one variableNano Banana 2
Brand flagship storeStore-wide visual systemMaintain an 8-image style layer long-term, rotating in new approved photosNano Banana 2
Batch text/background edits after the factRedoing images one by oneUse inpainting for uniform touch-upsNano Banana 2

Sign up on the Flux Art website and get 500 credits free. Tonight, turn 6 of your store's approved best-performing photos into a reference set; tomorrow, test-run 3 new SKUs. Plans and promotions are subject to the official site's current terms.

  • Flux Art official website (features, plans, and terms subject to change — check the current version)
  • Cyberspace Administration of China, Measures for Labeling AI-Generated and Synthetic Content: https://www.cac.gov.cn/2025-03/14/c_1743654684782215.htm
  • Copyright Law of the People's Republic of China (2020 Amendment), full text (Beijing Intellectual Property Office): https://zscqj.beijing.gov.cn/zscqj/zwgk/flfg18/436481084/index.html

Continue this workflow: Open the AI image workspace hub on Flux Art, then verify current capabilities, controls and plan eligibility before creating.

Open the AI image workspace →

FAQ

Basics

Q: Is the reference-image mechanism just the old-school 'toss in a reference photo' trick?

A: Same underlying idea, but this is a structured version of it: tiered quotas, only approved photos allowed in, and continuous rotation — it upgrades ad hoc reference photos into a maintainable "style asset."

Q: Why can't prompts alone keep style consistent?

A: Text descriptions leave room for interpretation, so the model reads them slightly differently each time. Reference images turn style into an executable visual baseline, which removes that layer of randomness.

How-To

Q: How should I split the 14 image slots?

A: 6–8 for the style baseline, 4–6 for category representatives, 0–2 for new-product shots. Better to leave slots empty than fill them with weak photos — low-quality images pull the whole set off-style.

Q: What's the right rhythm for batch generation?

A: Three constants, one variable: fix the model, aspect ratio, and prompt template, and only swap the product name and selling points. Generate 3 control samples to calibrate first, then scale up, and review the whole batch together once it's done.

Q: What if I want to change the background color across the whole batch afterward?

A: Use inpainting to repaint the background on each image — don't regenerate the whole photo. The layout and subject stay untouched, so consistency isn't broken.

Q: How often should I maintain the reference set?

A: Review it once a quarter — retire old images whose performance has dropped, and rotate in newly approved, high-performing photos.

Tool Choice

Q: Compared to just locking a single template, what's the advantage of a reference set?

A: A template locks the layout; a reference set locks the overall feel — color tone, lighting, and negative space carry through, while each image still keeps its own compositional freedom instead of looking like it was stamped from the same mold.

Q: Can multiple stores in a store group share one reference set?

A: The style-baseline layer can be shared; the category and new-product layers should be built per store. If you hand a store off to a new operator, the reference set transfers with the rest of the store's assets.

Pricing

Q: Does using reference images cost extra?

A: Credits are charged per generation task — check the official credit table for current rates. The 500 sign-up credits are enough to run through a full "build the set, calibrate, scale up" test cycle.

Risk & Compliance

Q: Will the platform store my reference images or use them for other purposes?

A: Check the current version of Flux Art's privacy policy — we can't make that commitment on the platform's behalf here. If you're unsure, build your reference set from de-identified images.

Q: Can I use a competitor's viral photo as a reference?

A: Not recommended — keep your copyright chain clean. You can study and recreate the structure yourself, but don't feed someone else's photo directly into the tool.

Q: Do AI-generated series images need to be labeled?

A: Yes — follow the Measures for Labeling AI-Generated and Synthetic Content. When publishing in bulk, make labeling a default step in your workflow.

Troubleshooting

Q: What if a newly generated image doesn't match the baseline's color tone?

A: First check the reference set's purity (remove one or two off-style images), then check whether you switched models or aspect ratios. Those two steps fix most drift.

Q: What if subject fidelity breaks down (color or material changes)?

A: Add a sharper subject photo to the new-product layer, and emphasize "keep the product's appearance unchanged" in the prompt. Compare each image against the actual product before listing.