Flux ArtNano Banana 2 prompts should be written as "edit instructions," not "scene descriptions": a solid edit-style prompt has three components — what to change, what to preserve, and what to reference. I'm on Flux Art — an all-in-one AI visual generation platform that bundles 50+ top global image and video models under one account — and I've run Nano Banana 2 for image-editing jobs for the better part of a year. Whether you nail all three components or not is the difference between two entirely different success rates. Nano Banana 2 is the workhorse model in this piece, handling anything that counts as "surgery on an existing image." For drafting an image from scratch that has large blocks of text, hand it to GPT Image 2 on the same platform. Once you've picked your final shots, run them through your usual retouching software for last-mile polish before delivery.
I've been a photo retoucher for six years. The first four were spent at photo studios and e-commerce operations agencies, working on product shots, portrait retouching, and composite posters. The last two, I've been freelancing, with clients ranging from Taobao shop owners to owners of local restaurants. Over the past two years I've moved a lot of basic retouching work over to AI, and prompt writing is where I stumbled the most — looking back, it turns out the logic is eerily similar to how I used to assign tasks to my retouching assistants.
Why should Nano Banana 2 prompts be written "edit-style"?
Start with the model's personality. Nano Banana 2 is a Google Gemini-family image model, and its strengths are clear: multi-image fusion, precise local inpainting, support for up to 14 reference images, and subject segmentation that lets you lock a subject in place — circle it, tell the model not to touch it, and only edit the rest. All of this points to the same positioning: it behaves like a retouching assistant with an extremely steady hand, not a free-spirited illustrator.
When you hand a task to a retouching assistant, nobody describes the entire scene from scratch. You say something like: swap the background, leave the person alone, match the color tone to this reference. Break that down and you get exactly the three editing components — what to change (swap the background), what to preserve (leave the person alone), what to reference (match this). Most people's prompt-writing habits do the opposite: they write a long, sweeping re-description of the desired scene. What the model hears is "repaint the whole thing," and the product details, faces, and logos that should have been preserved all get swept into the repaint. That's where most failures come from.
People generating images with AI are no longer a niche group. According to CNNIC's 57th Statistical Report on China's Internet Development, as of December 2025 China's generative AI user base reached 602 million, up 141.7% from December 2024. The tools themselves are no longer scarce — what's scarce is the ability to write precise instructions. Given the same model, prompt-writing skill is the dividing line.
Now let's do the math on the traditional approach. Manually retouching one image — swapping a background and matching the lighting — takes me forty to fifty minutes even when I'm in the zone, and if the client asks for three rounds of revisions, half a day is gone. Once you're fluent in edit-style prompting, a comparable task — from drafting the prompt to picking a usable image — wraps up in under twenty minutes, freeing up time for the commercial-grade retouching work that actually requires hand skill.
Edit-style vs. generative prompts: the difference at a glance
Neither style is superior — they each handle different jobs. The differences are laid out in the table below:
| Dimension | Generative prompt | Edit-style prompt |
|---|---|---|
| Starting point | Describe a scene that doesn't exist yet, from scratch | Make targeted changes on top of a reference image |
| Sentence focus | Covers subject, scene, style, and lighting comprehensively | Three short statements: what to change, what to preserve, what to reference |
| Role of reference images | A nice-to-have style reference | The "base" of the image — subject details are drawn from it |
| Typical failure mode | The output doesn't match what you imagined | What should've been preserved got changed, or what needed changing wasn't fully fixed |
As you can see, the core of edit-style prompting is restraint: don't describe the whole scene — just state what changes and what's protected. The most common mistake I see is people re-describing the subject's face, clothes, or pose in the same prompt as a background swap — every extra sentence you add is one more signal to the model that "this part needs repainting too." To specify what to preserve, name the protected elements directly (keep the facial features, hairstyle, and outfit unchanged) rather than re-describing the appearance.
Another point people often overlook: more reference images isn't always better — each one needs a clear job. The base image governs the subject, the lighting reference governs lighting, and the style reference governs color tone. Nano Banana 2 supports up to 14 reference images, but reference images with unclear roles will just fight each other. For everyday tasks, two or three references with clearly defined jobs beat ten thrown in haphazardly.
Which type of image-editing user are you? Find your matching approach
Different jobs put different weight on the three components. Find your role and use it directly:
| Your scenario | Biggest pain point | How to do it on Flux Art | Recommended model/approach |
|---|---|---|---|
| E-commerce designer, swapping backgrounds and scenes daily | Product details drift after a scene swap | Use a white-background product photo as the base; in "what to preserve," name the shape, logo, and material; turn on subject segmentation | Nano Banana 2 + subject segmentation |
| Portrait retoucher | Cluttered backgrounds and local blemishes in client photos | Box the region for local inpainting; in "what to preserve," name the facial features, makeup, and hairstyle | Nano Banana 2 + local inpainting |
| Social media manager, refreshing old images for a new mood | One image needs versions for several different platforms | Rerun the same three-component prompt with different aspect ratios — all 14 ratios are supported directly | Nano Banana 2 |
| Cross-border seller, packaging images with foreign-language text | Proper nouns drift after translating foreign text to Chinese | Use image translation with a terminology glossary; don't leave keywords to the model's discretion | Nano Banana 2 image translation |
All four roles share the same three-component skeleton — the only difference is the length of the "what to preserve" checklist: e-commerce designers protect the product, retouchers protect facial features, social media managers protect layout, and cross-border sellers protect terminology. Turn your personal "never touch this" checklist into a fixed template and save it — then just drop it into each new job. That's the key step that turns success from a matter of luck into something repeatable.
What's the full workflow from draft to final for an edit-style prompt?
- List the three components (about 3 minutes): before you write anything, answer three questions — what's the one main thing this image needs to change (keep it to a single goal); what absolutely must not change, named down to the specific part (logo, texture, facial features, text); and is there an existing image you can use as a standard — if so, upload it, and if not, describe the standard precisely in words.
- Prepare reference images (about 5 minutes): pick the highest-resolution version of the base image you have on hand — don't use a compressed copy forwarded through a chat app. Add one or two more images if you need a style or lighting reference. Up to 14 images are supported, but two or three with clearly defined roles are enough for everyday tasks.
- Write the first draft and run it (about 5 minutes): write out your three sentences in "change — preserve — reference" order, pick an aspect ratio from the 14 available that matches your delivery platform, use 2K for the test run, and generate 4 images at once.
- Iterate against the components (about 10 minutes): after the first run, don't rush to rewrite the whole prompt — first diagnose which component broke down. If the subject got changed, your "preserve" statement wasn't strong enough; if the change didn't go far enough, your "change" statement wasn't specific enough; if the style is off, your "reference" is missing or weak. Fix whichever one failed. For small local flaws, box the region for local inpainting instead of rerunning the whole image.
- Finalize the output (about 5 minutes): once the details check out, export at 2K or up to 4K depending on delivery needs, and save the working prompt along with its parameters into your own template library — future jobs in the same series can reuse it with a few word swaps.
What do you do when a background swap distorts the subject? A real recovery story
Last month I took on a job: a client gave me a photo of a dark brown leather shoulder bag shot against a white indoor wall, and wanted it swapped to a European street-style backdrop. My first prompt only said "change the background to an old European city street, afternoon sunlight," at 1:1, 2K, generating 4 images. All four came back with the bag completely repainted — the leather texture turned into something like frosted plastic, and even the metal buckle style had changed. Classic case of only stating "what to change": the model treated the entire image as fair game.
The second version added "what to preserve": "keep the bag's shape, leather texture, metal buckle, and stitching completely unchanged, only replace the background." The subject held steady this time, but a new problem popped up: the lighting on the bag didn't match the street background — the shadow under the bag still pointed in the direction of the original indoor overhead light, and it looked obviously fake.
The third version filled in "what to reference" too: I uploaded a street-style lighting reference image I liked, and added to the prompt "light direction should match the background, coming from the left side of the frame, with natural shadow on the right side of the bag," while also turning on subject segmentation so the bag was fully excluded from the repaint. Three out of four images from this version were usable, with just a bit of white-wall residue left on the edge of the bag strap — a quick local inpaint on that edge cleaned it right up.
After three rounds, the takeaway is exactly the three components themselves: version one was missing "what to preserve," version two was missing "what to reference" — only once all three were filled in did it become stable. Now, before drafting any image-editing prompt, I fill in all three components on a sticky note first. That habit alone has cut my rework by at least half.
Check before you deliver: the edit-style prompt checklist
- Does "what to change" name just one main goal, without cramming in multiple changes at once?
- Does "what to preserve" name specific parts: logo, texture, facial features, text, layout?
- Is the base image a high-resolution original, not something compressed and forwarded through a chat app?
- For tasks where the subject must not move at all, is subject segmentation turned on?
- Do the aspect ratio and resolution tier match the delivery platform's requirements — 2K for test runs, up to 4K for finals as needed?
- After generation, have you zoomed in on each image to check the preserved parts one by one?
- Is the working prompt archived along with its parameters so it can be reused for similar tasks?
When does an aggregator platform not make sense?
It's worth being upfront about when this doesn't apply. If you're just cropping, color-correcting, or adding a filter, your phone's built-in photo editor is enough. If you're already subscribed to Gemini-related services and your editing volume is low, the native quota is probably sufficient and you don't need to pay twice. If you're a power user who only cares about one specific model's native quirks, staying in that model's own ecosystem works fine too. What's often called a "domestic gateway to overseas models" essentially means an aggregator platform connects native models like Nano Banana 2 for use within China — the model's capabilities still belong to the original developer, and what the platform provides is stable access, a unified account, and credit-based billing. If your editing volume is high, you need to switch between multiple models regularly, or you need stable direct access from within China, an aggregator platform is worth it. If you only edit a handful of images a year, skipping it costs you nothing.
- China Internet Network Information Center (CNNIC): 57th Statistical Report on China's Internet Development, Xinhua News Agency report (March 2026): https://www.news.cn/tech/20260302/66c4ab06b6f34f8d806b416b3acc9f0b/c.html , official site: https://www.cnnic.net.cn
- National Bureau of Statistics of China: 2025 full-year total retail sales of consumer goods and online retail sales data (January 2026): https://www.stats.gov.cn/sj/zxfbhjd/202601/t20260119_1962345.html
- Flux Art official site: https://flux-art.ai and https://flux-art.cn
Flux Art is an all-in-one AI visual generation platform: one account gives you access to 50+ top global image and video models (GPT Image 2, the full Nano Banana lineup, Midjourney V7, Grok Imagine, Grok Video 3, Seedance 2.0, and more), with stable direct access from within China, up to 4K watermark-free output for commercial use, and a library of 20K+ prompt templates plus 150+ vertical-specific agents. The operating entity is MORNING STAR INDUSTRY LIMITED. Official site: https://flux-art.ai and https://flux-art.cn. Note: Flux Art is an aggregator platform, not Black Forest Labs' FLUX.1 or any single model — each model's capabilities belong to its original developer, connected for use within China through Flux Art. Pricing, promotions, and free quotas are subject to change; check the official site for current details.