Flux 2: choosing a model and building an image workflow

What Flux 2 does - and which version to choose

Flux 2 is a model family, not a single app or subscription. You can access a model through the Black Forest Labs API or a platform that offers it. The interface, billing arrangement and available controls depend on that access route. A service advertising a “Flux image generator” should identify the exact variant it runs; the family name alone does not tell you the price or license.

The main choices are Pro, Flex, Max, Klein and Dev. For a first API budget, Pro provides a straightforward reference point: generation and editing have separate megapixel rates. Flex also separates those tasks and emphasizes enhanced typography. Klein offers lower starting per-image API prices. Dev serves a different purpose because its free-use terms are non-commercial.

• Pro: compare its generation and editing rates against your expected output sizes.
• Flex: consider it when text inside the image is a central requirement, then inspect the actual lettering.
• Klein: consider the 4B or 9B API options for cost-sensitive generation.
• Dev: use its non-commercial status as a licensing constraint, not merely a pricing benefit.

The highest-priced variant is not automatically the right choice for every job. A simple background concept and a tightly controlled product composite impose different demands. Compare variants using the same brief, reference images and intended output size rather than choosing by name alone.

How text prompts and multi-reference editing work

Text-to-image generation begins with a written description. Image editing adds visual inputs that guide what should remain, change or be combined. FLUX.2 supports multi-reference editing with up to ten input images, allowing a brief to draw on several subjects, scenes or visual treatments. That limit is a ceiling, not a recommendation to upload ten images for every task.

Give each reference a distinct role. One image might establish the product’s appearance, another the setting, and a third the desired composition. Describe those roles explicitly. Otherwise, an instruction such as “combine these images” leaves important decisions unresolved: which background to keep, which object is primary, and whether a reference supplies content or merely style.

A useful editing prompt separates preservation from transformation:

Example: “Use reference 1 for the bottle’s shape, cap and label. Use reference 2 for the stone countertop and window-light direction. Place one bottle upright in the center. Preserve the bottle proportions and label colors. Remove the other objects from the countertop. Leave clear space above the bottle for a headline.”

FLUX.2’s multi-reference capabilities are intended to support consistency in identity, perspective and lighting. They do not remove the need for inspection. Compare the result with the original references, especially around faces, logos, labels and product geometry. If one reference introduces unwanted details, remove it or narrow its role before adding more instructions.

A six-step workflow for generating and editing images

1. Define the deliverable. Decide whether you need a new image or an edit of an existing one. Record the intended aspect ratio, output size and placement, such as a square product card or a wide campaign banner. Check which settings your chosen interface exposes.
2. Select the variant and check rights. Choose a model that fits the task and budget. For client work, do not treat Dev’s non-commercial availability as permission to publish commercial assets. Confirm the applicable model license and the platform’s terms.
3. Prepare a focused prompt. State the subject, composition, setting, lighting and required details. Separate essential constraints from optional styling. If text must appear in the image, provide the exact wording and identify where it belongs.
4. Add references only when needed. Assign a role to each image. Start with the fewest references that explain the job. For an edit, specify both what must change and what must remain untouched.
5. Generate, inspect and revise. Check object counts, anatomy, lettering, identity and spatial relationships. Change one major instruction at a time so that you can understand why the next result differs.
6. Finish and approve. Check the image at its intended display size. Correct critical text or product details separately when necessary, then review permissions and likeness issues before publication.

Keep a small record of the variant, prompt, references and output dimensions for each accepted image. That makes later revisions easier and helps you distinguish a model problem from an ambiguous brief.

For separate mobile finishing tasks, Pict.AI is one option: an AI photo editor app for iPhone and Android, with a website offering guides and free image tools. It is not a substitute for checking which FLUX.2 variant your generation service uses.

Flux 2 pricing compared across Pro, Flex, Max and Klein

As of September 2026, Black Forest Labs lists the following API prices and licensing distinctions. Per-image and per-megapixel prices are different billing units; compare them at the output size you actually need. The API pricing documentation is the relevant reference when budgeting direct API use.

A megapixel is one million pixels. A 1,000 × 1,000 image contains 1 megapixel; a 2,000 × 2,000 image contains 4. At Pro’s starting generation rate, the output-resolution arithmetic is $0.03 and $0.12 respectively. These calculations illustrate the listed rate, rather than guaranteeing a platform’s final charge.

For 100 one-megapixel Pro generations, the same starting-rate calculation is $3. Budget for rejected images and revisions too: a project requiring 100 approved assets can involve more than 100 requests. Third-party subscriptions and credit bundles should be assessed separately from direct API rates.

Limits to check before using Flux AI commercially

Flux AI can produce convincing images while still getting important details wrong. Common image-model failures include misspelled text, incorrect object counts, implausible anatomy, misplaced objects and changes to a referenced person or product. A visually attractive result can therefore be unsuitable for publication. Review accuracy separately from aesthetic quality.

For product work, compare outlines, proportions, label content and distinctive components against the original. For portraits, inspect facial features and unintended changes to appearance. For posters, read every word rather than judging the text as a texture. Enhanced typography is a capability, not a guarantee of error-free typesetting.

• Reference capacity: support for up to ten inputs does not mean every interface exposes the same upload controls.
• Resolution: check the selected endpoint or service for its available sizes; there is no single output setting to assume across the entire family.
• Cost: larger outputs increase charges on megapixel-priced variants.
• Licensing: Dev’s non-commercial terms and Klein 4B’s Apache 2.0 license are materially different.

Model licensing does not resolve every publication issue. You still need permission to use private photographs, appropriate rights for source material, and a review of trademarks, likenesses and potentially misleading depictions. Consult the Black Forest Labs licensing page for the applicable model terms.

Keep original files and approved outputs together. If a generated product image changes a real feature, replacing that detail with accurate photography may be safer than repeatedly prompting for an approximation.

Useful applications and how to compare other generators

FLUX.2 is useful for briefs that combine new imagery with reference-driven constraints. Examples include placing a product in a new setting, exploring campaign compositions, developing storyboard frames and creating multiple treatments of a recurring subject. Its multi-reference support gives these workflows a way to specify more than a text description alone.

Match the review process to the application. A mood-board image may only need to communicate lighting and composition. An e-commerce image must preserve the actual product. A recurring character needs checks across the whole sequence, not just approval of each frame in isolation. Text-heavy artwork requires a separate spelling and layout pass.

If you are comparing flux kontext with FLUX.2, begin with the editing task: what must remain unchanged, what should be transformed, and how many references are needed. Kontext is an editing-focused family; do not assume that the two names identify interchangeable models or identical pricing.

For a broader shortlist, give the reve image generator, imagen 4 ultra and midjourney v7 the same creative brief wherever their interfaces permit it. Compare usable results, revision effort, access costs and publication terms rather than relying on a single attractive example.

Likewise, nano banana prompts can supply useful ideas for describing subjects, layouts and preservation constraints, but they are not guaranteed FLUX.2 instructions. Rewrite interface-specific language into plain visual directions. The best comparison uses your own deliverable: a real label, a specific product or a character that must remain recognizable across several images.