Failure modes & post-production combos

99% isn't 100%.

After the first 16 chapters, you can crank out images reliably — posters, Xiaohongshu, WeChat, LinkedIn, logos, IP characters, infographics. The playbook's all there. But run gpt-image-2 on a real production line for a week and you'll find that lingering 1-5% failure rate clusters in four spots. This chapter is your first-aid manual. Each of the four failure modes gets a counter-prompt plus a rescue workflow, plus three post-production combos to take you from "can ship images" to "can deliver reliably."

---

1. Failure Mode 1: Hands and Anatomy Errors

Five fingers sometimes become six. Occasionally you get "two right hands" or fingers that fuse together. Way better than Midjourney, but it still happens — close-up shots of people screw up roughly 1 in 10 times. Far-shot crowd scenes fail more often, but the pixels are tiny so nobody notices. You can mostly let those slide.

Prevention prompt: For close-ups, lock this into your constraint block:

anatomically correct hands, five fingers visible per hand,
no extra digits, no fused fingers, natural finger spacing

Rescue workflow A: Photopea local inpaint. Box-select the broken hand area, run Photopea's built-in AI repaint (free). Two or three passes usually fixes it.

Rescue workflow B: Send it back to ChatGPT for a multi-turn edit:

Keep everything else exactly the same.
Only fix the right hand to have 5 normal fingers,
proper anatomy, no extra digits.

The model holds everything else, only touches the hand. This trick is rock-solid for single hands.

---

2. Failure Mode 2: Garbled Text

Chinese is 99%, not 100%. Short headlines (4-8 characters) almost never break, but headlines over 12 characters and complex stroke characters ("嬴", "鬱", "龘") occasionally come out wrong — looks fine at a glance, zoom to 200% and you spot a missing stroke.

Prevention prompt: Double quotes plus role hint plus exact text only — the trio (covered in Ch 06). After every high-stakes poster, zoom to 200% and proofread character by character. No skipping.

Rescue workflow A: ChatGPT multi-turn fixing single characters:

Keep everything the same. Only fix the headline:
the second character should be "学" not "字".
Use exactly the same font style and color.

Rescue workflow B (the most reliable safety net): Figma or Photopea typed text layer overlay. When generating, force no text, no overlay, leave headline area clean. Then add the text layer manually with a professional font (Source Han Serif / Alibaba Sans / Huawei HarmonyOS Sans).

Why is this safety net the most reliable? Because you bypass AI text rendering entirely — error rate drops from 1% to 0. Think of it this way: instead of sweating every image, let AI focus on the background and let font files handle anything that has to be 100% correct. Each tool does its job. Nobody steps on anybody's toes.

---

3. Failure Mode 3: Copyright and Sensitive-Word Blocks

Disney, Marvel, Pokemon, well-known public figures, politically sensitive terms — all blocked. The usual error is the model flat-out refusing, or quietly "blurring" brand features — you wanted Mickey Mouse, you got a grey mouse that kinda looks like him but kinda doesn't.

Counter-move: Replace brand names with generic descriptions.

For public figures, always use occupation + nationality + age range. Don't write the name. You still get the vibe you wanted, and you skip the block.

---

4. Failure Mode 4: NSFW Blocks

OpenAI's NSFW standard is strict. How strict? Even "a woman in a black fitted dress at a cocktail party" — perfectly normal business copy — can get blocked. Fashion accounts, yoga courses, and fitness posters trip on this constantly.

Counter-move: Conservative wording for clothing.

Yes — casual wear / professional attire / business casual
Yes — wearing a knee-length dress / wearing a fitted blazer
No — tight / revealing / sexy / form-fitting / low-cut

Domestic Chinese platforms are even more conservative than international ones. Don't go anywhere near the borderline. We learned this the hard way doing Xiaohongshu work: instead of arguing with OpenAI inside the prompt, just steer the scene toward neutral territory like "office / commute / meeting". The vibe lands the same way and your block rate drops to zero.

There's also a hidden trap: get blocked 3 times in a row and OpenAI temporarily rate-limits your account — for the next 1-2 hours every prompt comes back slow with a higher block rate. So the moment a prompt gets blocked, pivot and rewrite immediately. Don't keep slamming the same prompt over and over.

---

5. Post-Production Combo Workflows

You've got counter-moves for all four failure modes now. But on a production line you also need an actual workflow — generation, repair, and layout strung into one assembly line. Here are three independent workflows. Pick whichever fits your scenario.

Workflow A: AI Base Layer + Manual Text Overlay (Most Reliable for Chinese Posters)

1. gpt-image-2 generates the base
   prompt forces "no text, no overlay, leave [area] clean"
2. Download into Figma / Photopea
3. Add text layer using a professional font
   (Source Han Serif / Alibaba Sans / Huawei HarmonyOS)
4. Tune kerning / leading / color / shadow
5. Export

You skip the text-failure risk and you're 10x faster than designing from scratch in Photoshop. AI base layer in 30 seconds, manual typesetting in 3-5 minutes — single poster done in 6 minutes. For Chinese-heavy posters (event KVs, e-commerce hero images, course covers), this is the go-to.

Workflow B: AI Output + Photopea Inpaint for Local Repair

1. gpt-image-2 outputs (whole image looks good,
   only hands or single character broken)
2. Photopea selects the broken region
3. Photopea's built-in AI repaint, or send back to ChatGPT
4. Export

Good for "99-point image, off by 1 thing" situations. Hand wrong on a portrait, single Chinese character missing a stroke — regenerating the whole thing costs too much, local repair is the smart move.

Workflow C: AI Output + Canva Template Layout

1. gpt-image-2 generates the base
2. Drop into a ready-made Canva template
   (title / subtitle / logo zones already planned)
3. Swap the template's default colors and font sizes
4. Export

For non-designers who just need to ship. Canva templates lock down "what text goes where, how big" so you only worry about generating the base and tweaking copy.

---

6. Tool Cheat Sheet

No paid tool? Photopea plus free-tier Figma covers 90% of scenarios.

---

7. Pitfall Cheat Sheet

Print this and stick it next to your monitor. Glance at it before generating — your failure rate drops by half.

---

8. JR Academy: Our Experience

The JR team's first run with gpt-image-2 was the spring enrollment campaign. 800+ images in 4 weeks — posters, Xiaohongshu covers, WeChat headers, LinkedIn visuals, all mixed. Week 1 failure rate was 23%. One in four images needed rescuing. Wrong text, broken hands, off proportions, style drift — those four were the biggest offenders.

Week 2 we got systematic. Wrote down counter-templates and rescue workflows for each failure mode. Dumped them into a shared Notion page for the ops team. Week 3 we introduced the "200% zoom recheck" step, and complex Chinese characters defaulted to Workflow A (manual text layer). Week 4 the failure rate was down to 6%.

The key wasn't the model getting stronger — it was the workflow getting systematic. Same gpt-image-2, week 1 vs week 4, the entire difference was knowing what to do after a failure. It's like cooking. Same pan, same ingredients. Beginners screw up because they haven't built the emergency reflex for "too much salt, what now" or "burnt the pot, now what." Veterans aren't error-free. They just know the next move the second something goes sideways.

---

9. One-Line Takeaway

AI generates 80% + post-production 20% = a usable product. Publishing pure one-shot AI images straight to the feed? Readers spot it in a glance — and your account gets tagged "AI slop."

Treat "AI generates 80% + post-production 20%" as your moat. Wrong hand? Photopea. Wrong text? Figma typing layer. Copyright block? Generic description. NSFW block? Conservative clothing. Once every failure mode has a fallback, your output finally has a "reliable delivery" floor.

---

10. Wrapping Up the Whole /learn/ai-image Direction

Congrats on finishing all 17 chapters.

Your next move isn't reading more tutorials — it's getting your hands dirty. Pick the platform you know best (Xiaohongshu / WeChat / LinkedIn), follow the templates from the matching chapter, and ship 30 images. Then run every image through this chapter's four-failure checklist.

Come back in a week. Your visual output will at least 3x.

Design used to be the bottleneck. From now on it's the fastest, most controllable link in your entire content production chain.

---

Post-Production Enhancement Case Study

From awesome-gpt-image (CC BY 4.0). A real "Workflow D: phone-shot + AI enhance" case — faster than from-scratch generation, more original, less hassle.

Case: Pro-Grade Upgrade for an iPhone Photo (Image-to-Image Workflow)

Prompt:

Enhance this iPhone photo with ChatGPT so it looks like a professional photographer and designer worked on it.

This is the fourth workflow beyond the three in §5 — the image-to-image / enhancement mode:

1. You have an original photo (phone snapshot,
   old photo, or low-quality client-supplied material)
2. Upload to ChatGPT with a one-line prompt ("enhance this photo")
3. gpt-image-2 keeps the subject and composition,
   upgrades lighting / color / sharpness / detail
4. Export

The biggest win here: originality is higher than pure text-to-image (your photo is yours), it's fast (no need to write a prompt from scratch), and the failure rate is essentially zero — the skeleton comes from a real photo, AI just fills in atmosphere.

Good for:

• Old phone photos heading to WeChat / Xiaohongshu
• Client-supplied low-res material that needs to be "leveled up" for commercial use
• Your own product photos getting upgraded to commercial-photography quality
• Bulk upgrading WeChat Moments / IG grid content

Heads up: this falls under edit mode, doesn't burn a "fresh generation" quota, and the model dodges hand and text failures automatically — those elements are already "correct" in the source, so AI doesn't invent anything. A very steady safety net.

Created by @dezainaz_ceo · Featured in awesome-gpt-image