WeChat MP banner & inline images

You open the WeChat MP editor to publish an article.

The exhausting part has never been writing. It's the banner.

You've already burned 20 minutes scrolling Unsplash and every image that fits has been used by someone else. You settle for Pexels, another 10 minutes gone. In the end you give up and stitch something together in Canva, another 20 minutes.

WeChat MP runs one article a day, which means 40 minutes a day on banner and inline images. 365 days a year times 40 minutes equals 240 hours — that's 20 hours a month eaten by low-value work.

After gpt-image-2 this changed. This chapter walks through how to use it for a complete WeChat Official Account visual system: banner (900×383) + inline images (recommended 900×500) + 6 images that hold a consistent tone + text-heavy layouts like infographics and flowcharts.

---

1. WeChat MP Cover Sizing (the detail most people miss)

WeChat MP banners are pickier than Xiaohongshu's, because they have to obey two constraints:

1. Feed thumbnail ratio

• 2.35:1 horizontal (standard 900×383) — feed thumbnail
• 16:9 inline image (recommend 900×500) — paragraph image
• Square 4:3 isn't the norm anymore, kept for legacy compatibility

2. The platform auto-adds the article title underneath

This is the biggest difference between WeChat MP and Xiaohongshu:

The platform overlays a white-background, black-text article title underneath your banner.
If your image already has big text on it, it'll fight with the platform title and read like a "double headline".
Your banner prompt has to force bottom 30% must be clean / blurred / no text overlay.

A lot of people fall into this trap their first time using AI for a WeChat MP banner — the image looks great, then the platform title slaps on top and slices the whole thing in half.

---

2. Banner Prompt in Practice

The formula:

Wide horizontal banner 2.35:1 ratio (900×383),
[场景描述 with subject + lighting + style],
no text overlay,
leave bottom 30% clean for platform title placement,
modern editorial aesthetic.

Three keywords you can't skip:

• no text overlay (no text on the image)
• leave bottom clean (keep the bottom open)
• editorial banner (editorial banner, not a marketing poster)

Real example: an article about AI Coding

Wide horizontal banner 2.35:1, 900×383.

A developer's desk at night, three monitors showing code,
cyberpunk neon glow from screens, slightly cool color tone.
Hands lightly resting on a mechanical keyboard, no face shown.

No text, no overlay.
Leave bottom 30% completely clean (out of focus or dark) for title placement.

Style: cinematic, editorial photography, shallow depth of field,
moody but professional.

The output goes straight into WeChat MP, the platform stacks its title on top, and nothing fights.

---

3. Inline Image Rhythm — 5 images for one article, all on tone

A 2000-word article wants an image every 400-500 words, so you need 4-5 images that share a visual style.

The workflow:

Image 1 (banner): full prompt to set the tone, generate 4 and pick 1.

Images 2-5 (inline): use reference so gpt-image-2 holds the same style and outputs variants.

Continuing with the AI Coding article above:

Same style as the banner (cyberpunk dev desk, neon glow, cool tones),
change scene to:

[Image 2 — close-up] Cursor IDE close-up on screen,
hands typing on keyboard, no face.

[Image 3 — abstract] Terminal with green text,
abstract code-rain effect, very moody.

[Image 4 — comparison] Two monitors side-by-side,
left showing legacy Vim, right showing AI suggestions popup.

[Image 5 — concept] A robotic hand and human hand
both reaching toward the same keyboard, soft contact light.

Five images in 10 minutes, totally consistent in tone. Readers' eyes don't get pulled out of the article.

---

4. Text-Heavy Layouts: Infographic / Flowchart / Comparison Table

WeChat MP has a particular kind of inline image — infographics, flowcharts, comparison tables. These actively want text, and this is where gpt-image-2 really shines.

Infographic prompt template

Infographic, 16:9 banner, 900×500.

Title (top, bold): "ChatGPT vs Claude vs Gemini"

Three columns with headers:
- Left column header: "Speed"
- Middle column header: "Quality"
- Right column header: "Price"

Under each column, simple icon + one-line text:
- Speed column: lightning icon, "Sub-second response"
- Quality column: trophy icon, "Best at reasoning"
- Price column: dollar icon, "$2.50 / 1M tokens"

Style: minimal flat design, dark navy background, white + yellow accents.
No extra elements, exact text only.

This is the bread and butter of gpt-image-2 — 99% text accuracy pushes infographic usability from 30% in the MJ era up to 90%.

---

5. Where Things Blew Up

Blowup 1: bottom subject covered by the platform title

The first time I noticed gpt-image-2 packed the bottom too tight, the platform title (default white-on-black style) covered the entire bottom strip. For example, a desktop scene where the keyboard sits at the bottom 1/4 — platform title lands and the keyboard disappears.

Fix: force bottom 30% must be clean, blurred, or out of focus in the prompt. Or change the composition so the subject sits in the top 2/3.

Blowup 2: style drifts after 5 images in a row

Images 1-2 hold the tone, image 3 quietly starts to shift (cyberpunk slowly bleeding into vapor wave). The model "forgets" once the reference chain gets long.

Fix: re-paste the full style block every 2-3 images. Don't lean on the reference chain alone.

Blowup 3: 2.35:1 horizontal squashed back to 16:9

The ChatGPT web app sometimes outputs a 2.35:1 prompt as 16:9 (it's closer to 16:9).

Fix:

• Write absolute pixels 900×383
• Stack the hints wide horizontal banner, 2.35:1 ratio, ultrawide
• Hit the API directly and pass the size parameter

Blowup 4: infographic labels translated

Mixed a Chinese line and an English line in the same prompt and the model translated the English labels into Chinese (or the other way around).

Fix: whatever's inside the quotes is the final on-image text — don't let the model "interpret". Every label gets double quotes plus the full path.

---

6. What we learned: JR's WeChat MP @JR匠人学院

We switched to gpt-image-2 in early 2026-04. Four weeks later, the productivity numbers:

We saved 96 hours a month — basically all of one full-time operator's time from Tuesday through Friday.

But there's a hidden cost: style design moves up front. You have to decide the article's tone first (cyberpunk? editorial? minimal?), otherwise the AI has no idea which way to go.

What we did was preset a "prompt anchor" for each content column:

• "AI Tool Reviews" column: cyberpunk + editorial photography
• "Job Hunt Stories" column: warm light + real-people feel + cinematic
• "Tech Breakdowns" column: minimal flat + infographic style

Once the column's locked in, every article's 6 image prompts only swap out the subject description; the tone block gets reused.

---

7. API Cost Estimate

If you go API for batch production (worth it once you're shipping 50+ articles a month, or running multiple accounts):

For WeChat MP banners I'd go medium — the platform compresses the image anyway. Only the text-heavy infographics are worth high.

If you're on ChatGPT Plus / Pro you don't pay per image; the $20 / $200 subscription covers it.

---

What's Next

Next chapter is Moments — 1:1 squares + 9-grid composition narratives. Moments is your personal IP storefront, and the prompt style is completely different from WeChat MP.

If you're in a rush to swap out this week's WeChat MP banner:

1. Grab the horizontal banner formula from §2
2. Write a prompt with no text overlay + leave bottom clean
3. Generate 4, pick 1
4. Cross-check against the blowups in §5
5. Upload to WeChat MP and see whether it fights with the platform title

The single point that trips people up between WeChat MP and Xiaohongshu: Xiaohongshu wants text in your face, WeChat MP needs to step aside for the platform title. The prompt tone is completely flipped between the two.

---

📷 WeChat MP Real-World Test Cases

From awesome-gpt-image (CC BY 4.0). Real generation cases from the WeChat ecosystem.

Case: WeChat chat interface as a real smartphone photo (the self-media inline image)

Prompt (original Chinese):

生成一张图片，画幅 3:4。一张真实生活照风格的手机照片，手机屏幕上是中文对话的微信聊天界面，包含聊天气泡和 Word 文档附件。界面有绿色和灰色气泡。屏幕上有明显的指纹、污渍和划痕。玻璃有强反光和直射光产生的眩光。画面略微倾斜，手持拍摄，自然环境光，不完美的构图，强烈的真实感，随性日常的快照，高清细节，4K。对话内容是老板和员工：员工把文件发给老板，老板回复说先看一下。

If you write WeChat MP articles in the workplace / self-media / startup space, you constantly need a "fake WeChat chat screenshot" as an inline image. What's clever about this prompt is that it deliberately writes "imperfect" — fingerprints, smudges, scratches, glare, tilted framing — these "flaws" are exactly what makes the output look like a real phone snap instead of a composite. A counter-intuitive but effective realism trick.

📷 Creator: @Gorden_Sun · Collected in: awesome-gpt-image