FLUX 2 Pro Edit: A practical multi-reference editing guide

FLUX 2 Pro Edit is the production image-editing endpoint in the FLUX.2 family. It takes one or more reference images plus a natural-language instruction, then returns a newly rendered JPEG or PNG. The draw is straightforward: a designer can combine a product, material, setting, and visual direction without rebuilding the scene by hand.

That also sets the right expectation. FLUX 2 Pro Edit is a generative editor, not a Photoshop document with masks and editable layers. It can follow a precise instruction, but the output is still a newly generated image. Product teams should test identity, geometry, typography, and brand details before treating a result as final artwork.

What FLUX 2 Pro Edit accepts

The current FLUX.2 image-editing guide sends a text prompt with one or more ordered input images. The direct API accepts up to eight references, while the web playground accepts up to 10. Output can reach 4 megapixels.

The official FLUX.2 prompting guide adds an important capacity detail: the Pro API has a 9-megapixel total limit across inputs and output. At 1 megapixel of output, it can use up to eight references; at 2 megapixels, the documented limit falls to seven. "Use these images" still leaves the model to guess what each file contributes. A prompt that assigns a role to every reference is easier to test and revise.

For example:

Use image 1 as the exact product shape.
Use the brushed copper material from image 2.
Place the product in the dark studio from image 3.
Keep the camera angle and proportions from image 1.
Remove all existing labels. Return one clean 16:9 product image.

This is a better starting point than a long paragraph full of mood words. It separates the non-negotiable details from the visual direction.

A reliable multi-reference workflow

Start by deciding what each source controls. One image may define identity or product geometry. Another may provide a background, palette, material, pose, or composition. If two references disagree about the same property, say which one wins.

Then make the first instruction narrow. Change the setting while preserving the product. Once that passes, change the material. Add typography only after the composition is stable. Sequential editing costs more requests, but it makes failures easier to diagnose.

A practical FLUX 2 Pro Edit sequence looks like this:

1. Choose the strongest base image and crop away irrelevant material.
2. Add only the references needed for the current edit.
3. Name each reference and state the property it controls.
4. List what must stay unchanged, especially shape, camera angle, and subject identity.
5. Request one output before asking for variants.
6. Save the prompt, seed, source files, and returned image together.

The seed is useful for repeatability, but it is not a promise that every later edit will preserve every pixel. The source images and full request remain the real production record.

What current users are doing with FLUX.2 editing

A shared ComfyUI workflow shows the kind of demand behind FLUX.2 editing. The author uses the downloadable FLUX.2 dev variant to put a supplied character into a new setting, change clothing or eye color, and work with multiple references. This is community evidence about the workflow pattern, not proof that the Pro endpoint behaves identically.

Keep the variants separate. FLUX.2 dev is attractive when a team wants local control and configurable workflows. FLUX 2 Pro Edit is the hosted production option. FLUX.2 flex exposes more tuning controls and specializes in typography, while max targets the highest editing consistency.

How to estimate FLUX 2 Pro Edit cost

Black Forest Labs currently lists FLUX 2 Pro Edit at from $0.045 per image-editing request, while Pro text-to-image starts at $0.03. The official pricing documentation says FLUX.2 cost scales with output resolution, and the pricing help page says resolution rounds up to the next megapixel.

A 1,200 by 1,200 output contains 1.44 megapixels, so it is billed at the 2-megapixel tier before considering the reference-image portion of the edit. Reference count, source resolution, output resolution, and model variant can all change the charge. Use the current calculator for an exact estimate rather than multiplying one headline rate.

Storage, network transfer, retries, or the application supplying the model may add other charges. A production estimate should include the number of accepted images, the expected retry rate, and how many variants a reviewer normally needs.

Cost also changes with workflow design. Asking for four broad variations may be cheaper than making four serial corrections, but the broad set is harder to control. For a product catalog, a narrow, repeatable prompt often saves more reviewer time than it costs in extra calls.

Where FLUX 2 Pro Edit can fail

Multi-reference editing gives the model more context and more ways to get confused. Common failure categories include:

• the correct object with the wrong material or color;
• the right composition with altered product geometry;
• identity drift after several sequential edits;
• small lettering that looks plausible but is not correct;
• a reference style bleeding into an object it was not meant to control;
• a polished image that no longer matches the original camera or lighting.

Do not combine these into a single "quality" score. A visually attractive result can still be unusable for commerce if the product shape changes. Track hard failures such as incorrect text and geometry separately from subjective preferences such as mood or color grading.

Black Forest Labs reported a FLUX.2 Pro speed update in March 2026, describing roughly twice the generation speed with no price change. Speed helps iteration, but it does not remove the need for review. Faster wrong variants are still wrong.

A production evaluation that tells you something

Use 20 to 50 images from the work you actually ship. Include reflective products, faces, crowded scenes, tiny labels, unusual aspect ratios, and examples where the background color is close to the subject. Define the expected edit before anyone sees the output.

For every FLUX 2 Pro Edit result, record:

• whether the requested change happened;
• whether protected details stayed unchanged;
• whether text and logos remained accurate;
• how many calls were needed for acceptance;
• total review time alongside generation time;
• the final output size and cost.

The acceptance rate is more useful than a gallery of best results. It tells you whether FLUX 2 Pro Edit can support a repeatable catalog, campaign, or design workflow.

Using Yix alongside the editing workflow

Yix does not currently expose FLUX 2 Pro Edit as a selectable generation model. If you are starting from a reference image and need a structured instruction, the free Image to Prompt Generator can help you describe composition, lighting, materials, and camera language before you build the edit request. You can also browse the Yix image model directory for models that are available on the site now.

Analyze the reference first, then give the editor a short, role-based instruction. Keep the sources, prompt, seed, and final image together so a successful edit can be reproduced or audited later.

Related image-editing guide

For a different approach built around marked regions and editable layer output, read the Seedream 5.0 Pro Image Editing guide.

Sources: FLUX.2 image-editing guide, FLUX.2 model overview, official prompting guide, pricing documentation, Black Forest Labs release notes, and the community FLUX.2 edit workflow.

FLUX is a trademark of its respective owner. Yix is not affiliated with Black Forest Labs.