Blog

Guide

FLUX 3 Image Klein Prompt Iteration in Practice: From Product Composition to Multi-Reference Images

Using the fixed-variable method to gradually converge composition, reference image roles, and detail control

Nov 18, 2026
A geometric illustration of a mineral cross-section rendered in flat facets on a deep teal background, with clearly defined cyan-green and dark green facets, serving as the cover image for this article.
On this page

Why Product Visuals Need Staged Prompt Writing

The most common mistake in product visual prompts is treating them as an essay written all at once: subject, material, lighting, background, composition, camera, style, and negative descriptions are all stuffed into one paragraph, and then you generate, look at the result, feel dissatisfied, and rewrite the whole paragraph. The problem is that a single round of changes alters seven or eight factors at the same time, so you cannot tell which sentence actually made a difference, and the next round often just repeats the same luck in a different way.

The idea of staged iteration is to break it down into steps with sequential dependencies: first determine the frame and pixel tier, then determine the product's position and proportion in the image, then decide what role each reference image plays, and only finally refine details and the delivery layer. At each stage, only change the fields that belong to that stage and keep the others fixed. This is the only way the differences between rounds become explainable.

On the calling side, this method corresponds to a clear set of fields: prompt, image_urls, aspect_ratio, resolution, steps, seed, output_format, and safety_tolerance. Staging essentially means deciding which fields are allowed to change in this round.

The fixed-variable method has another layer of value: it turns prompts into versionable assets. Record the "prompt + parameter set + result" for each round. When a certain composition is adopted, you can trace back how it came about and reuse it when swapping in product images later.

Typical Failure Points of One-Shot Long Prompts

  • Piling on adjectives before the size is certain; after the composition stabilizes, you discover that the frame is completely unsuitable for the product detail page;
  • Throwing multiple reference images in together without specifying roles in the prompt, leaving the model to guess which image is the subject;
  • Mixing detail words (texture, highlights, material) with composition words in the same round of changes, so the lighting becomes clear but the layout goes off track;
  • Repeatedly adjusting delivery-layer parameters such as output_format and safety_tolerance during the exploration phase, creating irrelevant noise.
Stacked crystals conflict easily; split the variables apart and change only one per round, and the facets gradually become clear.

Set the Frame and Pixel Tier First, Then Discuss Image Content

Composition problems are often not caused by badly written prompts but by the wrong frame from the start. aspect_ratio determines the output frame, defaulting to auto, in which case the model decides the frame based on the prompt and reference images. Auto mode is convenient during exploration, but its judgments are unstable: with the same prompt, changing the word order slightly can produce a vertical or square image.

Once your target output is clear (homepage banner, horizontal detail-page image, vertical mobile feed image), you should fix aspect_ratio explicitly and stop letting the model decide for you. The available values cover horizontal, vertical, and square categories from 21:9, 16:9, 3:2, 1:1 to 4:5, 9:16, and more, enough for most shelf placements.

resolution is the pixel tier, with options 768, 1K, 1.5K, 2K, and 4K, defaulting to 1K. The specific width and height vary with the frame, so it forms a pair with aspect_ratio: the frame sets the ratio, and the tier sets the clarity. During composition exploration, it is advisable to stay at the default tier, because what you are judging at this point is layout and proportion, not edge sharpness; raising the resolution early only makes each round take longer and does not make composition judgments more accurate.

Mapping Frames to Business Placements

  • Horizontal images (such as 16:9, 3:2): suitable for homepage banners and horizontal scene images, with space on the left and right for copy;
  • Vertical images (such as 4:5, 9:16): suitable for mobile feeds and the first screen of detail pages, where the product is usually more stable centered and slightly lower;
  • Square images (1:1): suitable for shelf thumbnails and main images that need to be reused across channels.

Once these two are fixed, changing prompt will no longer lead you to mistake "wrong frame choice" for "ineffective prompt." When you need a quick check of field definitions, you can first look at the parameter descriptions on the model page.

Frames of different ratios are like facets of the same crystal: first choose the grid, then discuss what to put in it.

Describe Product Composition with Tags and bbox

The problem with using adjectives to describe composition is that they are not localizable. Sentences like "the product is centered, with space left above, and a soft shadow in the lower right" can be understood by the model, but the granularity of understanding varies each time. prompt supports combining <tags> with bbox JSON to specify layout and local editing regions, giving you a more "sketching" way to express things.

The basic approach is to name the key regions in the image and then use bbox to describe the positions of these tags. Tags handle semantics (this is the product, this is the white space, this is the shadow region), and bbox handles geometry (where its range is). This turns composition description from vague light-and-shadow imagination into a set of region definitions that are readable, editable, and comparable.

In practice, it is advisable to keep the number of tags restrained: the product subject, the white space where copy will be placed later, and the shadow/reflection area—usually three is enough. The more tags there are, the more they tend to squeeze each other, which is less direct than adjusting the frame.

Self-Check Questions for the Composition Stage

  • Does the product's proportion in the image meet the shelf requirements? Main images usually need to be fuller, while banners need more white space;
  • Is the white space actually usable for copy, rather than being half cut off by the product edge?
  • Do the shadow direction and the light source description match, avoiding shadows on both sides?
  • Does the background hierarchy interfere with the subject? If a clean background is needed, it should be clearly stated at this stage.

Before the composition is stable, do not rush to add material detail words. Detail words divert the model's attention to the surface and easily break up the region relationships you just adjusted.

The canvas is divided into localizable blocks: the main product area, the shadow area, and the copy area, like the facets of a mineral slice.

Assigning Roles to Multiple Reference Images: Subject, Style, and Composition

image_urls is an array of reference images, with image addresses as elements; omitting it means pure text-to-image. The most common misuse when multiple images are supported is "throwing in a few more never hurts": product photos, competitor style images, and layout references are all thrown in, while the prompt only says "generate a product image referencing these images." The model can only assign responsibilities on its own, and the result is often that the style image contaminates the product's appearance, or the layout image overrides the subject's proportions.

The correct approach is to state each image's role in the prompt. Roles usually fall into three categories: subject (which image determines the product's shape, structure, and color scheme), style (which image is referenced for lighting, texture, and atmosphere), and composition (which image is referenced for layout and white space). One image can serve multiple roles, but this must be clearly stated in the text; you cannot assume the model will understand.

In addition, prompt also carries the responsibility of "specifying what needs to be modified and what needs to be preserved in the reference images." That is, besides declaring who is who, you must also declare what to copy and what to change. For example, instructions such as preserving the product shape, replacing the background, and adjusting lighting should correspond one by one to specific reference image roles.

Writing Order for Role Assignment

  1. First, state in one sentence what the final image should be (for example, the main image for a certain product);
  2. Declare the role of each reference image one by one, linking them with tags;
  3. State the items to preserve and the items to modify;
  4. Finally, add overall atmosphere descriptions such as background and lighting.

There is an upper limit on the number of reference images, but what really determines success is not the count but whether each image has a clear responsibility. You can first use one subject image to get the composition working, then add a style image, and finally add a layout image, adding only one at a time and observing how the result changes.

The Fixed-Variable Iteration Method: Change Only One Thing per Round

To turn iteration into a process, there is only one core rule: change only one thing per round. Specifically, first lock a set of baseline parameters, then replace only one field in each round—either change one sentence in the prompt or add/remove one reference image.

For baseline parameters, it is recommended to lock seed, steps, and the frame pixels: take a fixed integer for seed, fix steps near the default upper limit to preserve detail space, and use the values determined in the previous step for aspect_ratio and resolution. After that, reuse this parameter set for each round of generation, and record the results by "round number, what was changed, and what the result was."

Have reasonable expectations for seed: its purpose is to "reproduce the same result as much as possible under the same parameters," not to guarantee consistency. Even if the parameters are completely unchanged, do not assume two outputs will be pixel-identical; treat it as a tool to reduce random variation, not as a promise of reproducibility. Therefore, the prompt version and the parameter set must be recorded together; otherwise, you cannot later determine whether a difference comes from the prompt or from randomness.

Record Template for a Single Iteration Round

  • Round number and goal (what you want to verify in this round);
  • Snapshot of fixed parameters: seed, steps, aspect_ratio, resolution;
  • The only variable in this round: which sentence in the prompt, or which reference image;
  • Result judgment: goal achieved / ineffective / introduced new problems;
  • Next step: keep, revert, or start a new branch.

When a round introduces new problems, the correct action is to revert to the previous parameter set and then change one other thing, rather than layering patches on the problematic version. When running multiple branches in parallel, use numbers to distinguish branches to avoid confusion in prompt versions.

Detail Refinement: steps, Resolution, and Local Editing

Only after the composition and reference image roles are stable do you enter the detail stage. At this point, the fields you can change are steps, resolution, and the local editing regions in the prompt.

steps is the number of diffusion steps, defaulting to 16, with a range of 1–16. Lower values generate faster and may reduce detail. Its role is a speed-versus-detail knob, not a quality-score knob: adjusting it before the composition is settled is meaningless, and after the composition is determined, use it to trade acceptable time for fuller texture and edges. Note that it does not affect the billing basis, so the trade-off criterion should be your waiting cost, not your budget.

resolution is only worth raising at this stage. The default 1K is enough to judge composition and role assignment; once the layout is finalized and you need to deliver a higher-pixel final product, raise it tier by tier and check whether over-sharpened materials or background noise appear at high pixel counts. Raise only one tier at a time for easy comparison.

Local editing relies on tags and bbox in the prompt: first use tags to enclose the region to be changed, then explain in the text what to change inside the region and what to preserve outside it. A common failure in local editing is that the region boundary does not match the product edge, causing a corner of the product to be cut off; therefore, when framing the selection, you can expand it slightly and then use text to constrain the preserved items.

Suggested Order for the Detail Stage

  1. First use steps to find the balance between detail and waiting time;
  2. Then use local editing to fix individual flaws, such as overly harsh shadows or edge color fringing;
  3. Finally raise resolution to produce the final image;
  4. Raise only one tier or change only one region at a time, confirming each item.

Delivery-Layer Parameters and Troubleshooting Common Issues

output_format and safety_tolerance are delivery-layer parameters. output_format defaults to jpeg and can be set to png; choose png when you need a transparent background or subsequent secondary processing, and jpeg is lighter for direct display on shelves. safety_tolerance is the content safety tolerance, an integer where 0 is the strictest, defaulting to 2. Neither should be repeatedly adjusted during the composition exploration phase: changing them early only introduces variables unrelated to the image. Determine them uniformly once the composition is stable and you enter batch production.

Common Issues and Troubleshooting Directions

  • The prompt changed but the image did not: first confirm whether only that one thing was changed in this round, then check whether the change falls at the level the model focuses on (composition words changed during the detail stage are often ineffective);
  • A reference image is ignored: check whether the prompt declares a clear role for that image, and whether the items to preserve and modify are stated;
  • Multiple images interfere with each other: reduce the number of reference images, add them one by one in the order of subject, style, and composition, and identify which one is causing trouble;
  • Reproduction fails: seed only reproduces as much as possible; confirm that the parameter set and prompt version are exactly the same, and accept that there will still be variation;
  • The frame does not match expectations: confirm whether aspect_ratio is still auto, fix it explicitly, and then judge the composition;
  • Generation becomes slow: check whether resolution or steps was raised too early, and return to the exploration-phase tier to verify.

Checking the definitions of the above fields in one place before running the next round usually saves a lot of ineffective attempts. For field details, see the flux-3-image-klein model page.

Staging, fixing variables, and recording each round—the value of this process is not to make the model smarter, but to make every change you make explainable. Treat prompts and parameter sets as code that needs version management, and product visual output will turn from guesswork into reproducible engineering.

Related articles

View all articles →

Start building with persistent AI artifacts

Get your API key and call image, video and audio models through one task-based API.

Read the API docs