Flux ArtYes, and mood and atmosphere are exactly where it shines; the weakness is just as clear — on-image text often comes out wrong, Chinese text especially so, and if you ask it to lay out the words on a poster directly, it fails more often than not. The workable approach is to split the job into two layers: hand the background to Midjourney V7 for atmosphere, then render the headline and info text with GPT Image 2 or add it in layout software afterward. On Flux Art — an all-in-one AI visual generation workspace that gives one account access to 50+ leading global image and video models — V7 and GPT Image 2 sit in the same workspace, so one poster can be produced by tag-teaming the two models, used directly from the web in China.
I've run event planning for four years — mall promotion cycles, store openings, weekend markets, all of it. Small-event material budgets are often painfully thin; a single outsourced poster can cost as much as half a month's materials budget, so since the year before last I've made posters myself with AI almost every time. This two-layer split was hammered out event by event.
Why does Midjourney's poster fall apart the moment you add text?
Here's the conclusion first: broken on-image text is a widely reported, publicly documented issue with Midjourney, and my own hands-on testing confirms it — English occasionally comes out right, but Chinese is almost always garbled, with strokes fused together, distorted shapes, and extra or missing strokes as the norm. There's no need to wait for a fix; treat it as a trait of the model and route around it in your workflow.
A poster, at its core, is really two layers: "image" plus "information." The image layer carries the emotion — whether a passerby stops to look depends on mood, color, and composition, and that layer happens to be exactly where V7 excels; its artistic, stylized output is widely recognized as strong. The information layer carries the conversion — the time, place, offer, and event rules. Get one character wrong there and the poster is essentially wasted. The two layers have completely different tolerance for error: a so-so image is just underwhelming, but wrong information is an incident.
Using AI for marketing materials stopped being a novelty a while ago. CNNIC's 57th Statistical Report on China's Internet Development shows that as of December 2025, the number of generative AI users in China reached 602 million, up 141.7% from December 2024. Among fellow event planners I talk to, nearly everyone is already using it — the difference is just whether they use it crudely or carefully.
Where's the care in the details? It's in the layering. I took my share of detours too: either relying entirely on template tools, which meant my poster looked like the shop next door's, or expecting a single model to handle image and text in one shot, which meant the text came out unusable and I was redoing work until midnight. Once you split the layers, the image has personality and the text has zero errors — you get both.
Who handles what on a poster? One table makes it clear
Break a poster into four parts, and each one gets its own owner:
| Poster component | Who handles it | Why |
|---|---|---|
| Mood background | Midjourney V7 | Strong artistic output — its specialty is mood and style |
| Headline and large text | GPT Image 2 | Accurate text rendering, 12 precision/resolution combos, up to 4K |
| Physical-object correction | Nano Banana 2 | Reference-image locking and inpainting keep in-store objects true to life |
| Small print and layout | Layout software | Zero-tolerance info like time, place, and phone number is safest set in real type |
Two notes worth unpacking. First, headline text has two paths: for short headlines where you want a stylized lettering effect, render it directly with GPT Image 2; for text-heavy info that needs repeated revisions, adding it in layout software afterward is far more efficient — revisions don't require regenerating the image. Second, if your store's actual physical elements appear on the poster — a sign, a product, a mascot — V7 will very likely give you a "reimagined" version instead of a faithful one. In that case use Nano Banana 2 with a reference photo to lock the shape, rather than fighting V7 into compliance.
What kind of event are you running? Match your scenario to a plan
Different events run into different poster pitfalls — find your scenario below:
| Your scenario | Biggest pain point | How to handle it on Flux Art | Recommended model/approach |
|---|---|---|---|
| Store opening or anniversary | Thin budget, no time to outsource | V7 generates a 3:4 portrait background, headline rendered with GPT Image 2, small print added afterward | V7 + GPT Image 2 |
| Mall promotion cycle | One key visual needs to extend to multiple sizes | Once the key visual is finalized, recompose it at different aspect ratios for landscape and portrait versions | V7 key visual + multi-ratio extension |
| Weekend market or pop-up | Style needs to feel fresh, no repeats each time | Swap prompt style each time — illustration, film, retro, in rotation | V7 + prompt template library |
| Online event banner | Heavy text, frequent revisions | Generate the background once, keep all text as a separate layer, revisions never touch the background | GPT Image 2 or V7 background + post-layout |
One line to sum it up: the deciding factor is "how much text, how many real objects." Heavy text relies on post-layout, lots of real objects rely on Nano Banana 2, and if both are minimal, let V7 run free.
What does the full workflow for a store-opening poster look like?
- Lock the copy first (about 10 minutes): finalize the headline, date, address, and offer details and proofread character by character — text is the zero-tolerance layer, don't wait until after the image is generated to edit it.
- Generate the background with V7 (about 15 minutes): use a 3:4 portrait ratio, write the prompt to cover only scene, lighting, and tone, explicitly leave out any text requirement, and add a line like "keep the upper third of the frame clean negative space"; try 4 images at the lowest tier first to test directions.
- Finalize and upscale (about 10 minutes): once you've picked a composition, rerun the same prompt at 2K for 2 backup options.
- Headline and info layer (about 15 minutes): render the headline with GPT Image 2 — use the background as reference, and spell out the exact text content and font feel; export the small print like date and address and add real type in layout software.
- Export and check (about 10 minutes): export at 4K with no watermark, produce print and online versions separately, and run through the checklist below item by item.
Chinese text turned into gibberish on a coffee shop's opening poster — how I fixed one real failure
Last month I made an opening poster for a friend's coffee shop and, wanting to save time, fell into a classic trap: I wrote straight into the Midjourney V7 prompt something like "a poster reading 'Grand Opening, 20% off storewide,'" at 3:4, lowest tier, 4 images. The image itself was flawless — warm light, a wooden bar counter, a close-up of latte art, mood nailed in one shot. But the "Chinese text" on all four images was unreadable, strokes smeared into a blob, like some alien script. That's not an operator error — broken on-image text is a well-documented, publicly known issue with this model. The fix took three steps. First, I stripped every text requirement out of the prompt and added "leave clean negative space at the top of the frame," reran 4 images, and picked the one with the steadiest bar-counter composition and the most complete top-margin space, then reran it at 2K for a high-res version. Second, I made the headline "Grand Opening" separately with GPT Image 2: uploaded the V7 background as a reference and wrote out the exact text content, its position (the blank area at the top), and the font feel (rounded handwriting style) — text rendering is its strength, and it took two rounds to get a usable version. Third, I didn't gamble on any model for the small print — date, address, discount details — and added real type in layout software after exporting. The day the finished poster went up, I made a point of standing outside the shop for a while — every character held up to a close look, and that's all that mattered; I'm not going to make up engagement numbers.
Check this list before you print or publish: the poster checklist
- Proofread character by character: headline, date, address, phone number, offer details — zoom to 100% and check every character.
- Clear information hierarchy: the headline should be readable from three meters away, and all key details readable from one meter.
- No leftover gibberish in the background: crop out or retouch any fake text V7 tossed into a corner of the background.
- Physical objects match reality: products, signage, and other real elements shown on the poster should match the real thing — inpaint locally if needed to fix them.
- Resolution meets spec: export at 4K for print; for large-format banners, confirm size requirements with your print vendor beforehand.
- No watermark: check all four corners and shadow areas on every exported version.
- Platform compliance: if you're publishing online and the platform requires an AI-generated content label, apply it per the platform's current rules.
When doesn't an aggregator platform make sense?
If your poster is purely text layout — white background, large type, a QR code — a template tool is enough, and there's no need to bring in a generative model; if you've already subscribed to Midjourney directly and your posters are mostly pure imagery or light on English text, your existing workflow can keep running as-is. And to be clear about how this all works: a "domestic gateway to overseas models" is, at its core, an aggregator platform connecting original models like Midjourney V7 for use within China — the model capability belongs to the original maker, and the platform provides stable access, a unified account, and credit-based billing. One more boundary worth flagging: extremely large-format banners. Generated images can show their limits when blown up to the extreme, so for important projects, build in time for a proof print — test a small sample before scaling up.
- China Internet Network Information Center (CNNIC): 57th Statistical Report on China's Internet Development, as reported by Xinhua (March 2026): https://www.news.cn/tech/20260302/66c4ab06b6f34f8d806b416b3acc9f0b/c.html , official site: https://www.cnnic.net.cn
- National Bureau of Statistics of China: full-year 2025 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 workspace: one account gives access to 50+ leading 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 direct, stable access in China, output up to 4K with no watermark and commercial-use rights, plus 20K+ prompt templates and 150+ vertical-specific agents. The operating entity is MORNING STAR INDUSTRY LIMITED. Official site: https://flux-art.ai and https://flux-art.cn. One clarification: Flux Art is an aggregator platform, not Black Forest Labs' FLUX.1 or any single model; each model's capability belongs to its original maker, made accessible in China through Flux Art. Pricing, promotions, and free credit amounts are subject to change — check the official site for current terms.