GPT Image 2, Nano Banana 2.1, FLUX 3 Image, Midjourney V8.2, and Seedream 5.0 Pro offer different ways to create and revise visuals. For developers making landing-page illustrations, onboarding images, or help graphics, the choice starts with what the asset must preserve. Selecting an API for a user-facing generation feature requires a separate integration evaluation.
A polished image can still be wrong for an application. The subject may sit behind the headline. Three onboarding illustrations may look unrelated. A diagram may contain an arrow that describes a transition your system never makes.
A rough draft gives you something concrete to revise. In Flux3, you can describe an illustration to generate a starting composition, or upload an existing image and request a change. Keep the brief and original image: they will help you compare candidates on the same assignment instead of judging unrelated showcase images.
Define the asset before choosing the model
Suppose you are building a task-management app. You need a landing-page illustration, three onboarding images, and a graphic explaining how a task moves through the system.
The landing-page image needs room for a headline. The onboarding set needs a consistent mascot and style. The explanatory graphic needs the right labels and relationships. These requirements lead to overlapping shortlists:
| Your hardest requirement | Candidates to compare | What narrows the choice |
|---|---|---|
| A hero illustration with reserved space for copy | GPT Image 2 and FLUX 3 Image | Start with a written composition brief. If you need to specify object regions explicitly, investigate FLUX's bounding-box workflow and whether your interface exposes it. |
| A consistent set built from approved references | Nano Banana 2.1, FLUX 3 Image, and Midjourney's Edit Model | Check whether the workflow accepts the references you actually need, then compare subject preservation across the full set. Input capacity alone does not establish consistency. |
| A visual style you have not settled on | Midjourney V8.2 alongside a candidate already on your shortlist | Personalization is a reason to investigate Midjourney. Compare treatments using the same subject and layout; other models can explore styles too. |
| An explanatory graphic with fixed labels and relationships | Seedream 5.0 Pro and GPT Image 2 | Compare the finished diagram against your source. Sketch-led or spatially annotated revisions are a reason to investigate Seedream; neither model's text capabilities establish factual correctness. |
If you need both a consistent mascot and precise placement, look for a workflow that accepts your reference and exposes the placement control you need. FLUX 3 Image supports both multi-reference input and bounding boxes, so it belongs on that shortlist. You still need to check the resulting mascot and layout together. The sections below explain reasons to investigate each model, rather than rank their output quality.
1. GPT Image 2: A detailed brief with several constraints
GPT Image 2 combines generation and editing with high-fidelity image inputs. OpenAI's guidance also covers text rendering and structured visuals, making it a candidate for both the hero illustration and the explanatory graphic.
For the hero, imagine three work cards arranged toward one completed stack. The cards should remain on the right, with quiet space on the left for a headline. A written brief can describe the objects, their relationship, palette, and empty region in one request.
If the scene is right but crowds the headline area, revise that existing image and check whether the edit preserves the cards and palette. This tests whether a text-led workflow gives you enough control for the layout. If regional placement remains the blocking requirement, compare it with the explicit box controls described in the FLUX section.
For a diagram, the test changes: preserving all requested objects is not enough if an arrow connects the wrong states. Use the diagram assignment below rather than choosing from a successful hero illustration.
2. Nano Banana 2.1: A workflow built around reference images
Nano Banana 2.1 supports conversational editing and up to 14 reference images, with output options at 1K, 2K, and 4K. Its reference workflow is relevant when you already have approved visual material.
For the onboarding set, supply an approved mascot, a palette reference, and a sample illustration. Give each input a role: preserve the mascot's shape, use the palette for color, and borrow the sample's rendering style. Then request separate scenes using that same reference set.
The 14-image allowance is useful only if those inputs resolve distinct requirements. With three references, capacity does not distinguish it from Midjourney's four-reference Edit Model. If your assignment genuinely needs more than four separate references in one request, that limit becomes a practical selection condition. It is not evidence that one model preserves details better.
Judge the three onboarding images together. A correct mascot in one image does not establish consistency across a set; changing proportions or line weight may require revisions even when each image looks good alone.
3. FLUX 3 Image: Composition you can specify explicitly
FLUX 3 Image offers bounding-box controls for placing elements and targeting edits, alongside written instructions and multiple reference images. Those controls matter when your design already says where things belong.
Suppose the hero reserves its left side for copy and its right side for the illustration. For a first draft, describe that arrangement in Flux3 and choose the aspect ratio used by your layout. If you already have an illustration, upload it and request a specific change, such as moving the stack right while preserving the cards and colors.
That browser workflow uses text and uploaded images. If the task requires explicit source and destination boxes, use an interface or integration that exposes the model's bounding-box controls. A prompt field and aspect-ratio selector alone do not provide that control.
A bounding box lets you name a region more explicitly than “slightly farther right.” It does not remove the need to inspect the surrounding image: Black Forest Labs notes that nearby shadows, reflections, or lighting may change after an edit. Compare those areas as well as the moved object.
4. Midjourney V8.2: A visual direction you still need to discover
Midjourney V8.2 focuses on aesthetics, image quality, and Personalization. Investigate Personalization when exploring a visual direction around your preferences, then judge whether that treatment fits the rest of the product.
For the task-management app, compare a paper-cut illustration, a restrained 3D scene, and a hand-drawn treatment. Keep the subject and intended placement stable. Otherwise, a clearer composition can make one treatment look preferable without telling you much about the style itself.
V8.2 also works with the Edit Model, which supports written changes, up to four reference images, and targeted editing or canvas expansion through the Editor. Include a revision in your trial: ask the chosen treatment to accommodate a changed object position without losing its look.
If the visual direction is already approved, style exploration is no longer your main selection criterion. Compare reference preservation and editing instead, using the same assets you would give Nano Banana or FLUX.
5. Seedream 5.0 Pro: A graphic that has to explain something
Seedream 5.0 Pro targets text-rich graphics and supports editing guided by spatial annotations and sketches. ByteDance also describes layer separation as part of its capabilities.
For the task-management graphic, a rough diagram can establish the order and grouping before you request a visual treatment. If a connection is wrong, marking the offending arrow on the source is a concrete revision method to investigate. Test whether those annotations make your corrections easier to specify and whether the result follows them.
Compare it with GPT Image 2 on the same labels, states, and connections. The technical meaning comes from your application: a polished extra arrow can make the graphic wrong, even if all the words are readable.
Choose the final format before investing in revisions. A static illustration may suit a help article, while a frequently changing process diagram may be easier to maintain as SVG or diagram source. Layer separation does not establish that the generated output contains editable text, semantic connections, or the application structure you need.
Compare the work left after generation
Choose the assignment that matches your asset, then give it to two plausible candidates. Keep the source images, required content, and target placement fixed. Adapt syntax and supported controls to each workflow; you are evaluating the way you would actually use it.
Hero illustration: composition and revision
Use this shared brief:
Create a wide illustration for a task-management app. Show three simple work cards arranged toward one completed stack. Keep the objects on the right, with an uncluttered area on the left for a headline added later. Use navy, cream, and one orange accent. Do not include text or logos.
Then request the same change: move the stack farther right while preserving the three cards and palette. Check that the headline area stays clear and that the revision does not replace details you wanted to keep.
If style is the open question, repeat this composition in the three treatments described above. Compare each at the same display size. Keep the actual headline and buttons in HTML so they remain editable.
Onboarding set: consistency across three scenes
Attach the same approved mascot, palette, and style reference to each candidate. Request three separate images: the mascot creating a task, assigning it to a teammate, and marking it complete. Keep the mascot's proportions, clothing, colors, and rendering style consistent.
Place the three results side by side inside your onboarding cards. Then change the pose in the second image while asking to preserve its established appearance. Compare both the original set and the revised set. This tests whether an individual correction creates new inconsistency elsewhere in the visual series.
Explanatory graphic: labels and relationships
Supply a simple source diagram with three nodes labeled “Queued,” “Assigned,” and “Complete.” Require exactly two directed arrows: Queued to Assigned, and Assigned to Complete. Ask for a navy-and-cream treatment without adding nodes or connections.
Check spelling, arrow direction, and legibility at the help article's display size. Then ask to change “Assigned” to “In progress” while preserving the other labels and both connections. If you want to evaluate annotation controls, mark that same node on the source rather than changing the requested result.
Record what it takes to reach an acceptable asset
Generate the same small number of candidates in each workflow and include the same revision task. Avoid choosing one fortunate output while ignoring the others. Record:
- Whether the initial asset met its required content, layout, or consistency conditions.
- Whether the revision preserved the parts that were already correct.
- The attempts and manual corrections needed to obtain an acceptable asset.
- Whether that asset still works inside its actual page or component, including its crop and display size.
Save the model version, settings, prompt, and references beside each result. Separate generation time from correction time. For cost, divide the total spent on the trial, including unsuccessful attempts, by the number of accepted assets. If none meet the requirements, record that outcome rather than treating a cheap generation as a usable result.
Start with the hardest requirement
For a layout problem, compare text-led revision with explicit placement controls. For a visual series, compare the full set and one revised member. For an explanatory graphic, compare correctness before polish. If style is undecided, hold the composition fixed while exploring treatments.
Choose the workflow that meets your non-negotiable requirements with an acceptable amount of correction. If both candidates fail the same requirement, reconsider the input or the asset format before testing three more models. Keep the accepted result and its revision record as a baseline when your project or model version changes.
Top comments (0)