Hello everyone, I'm Cang He.
A few days ago, GPT-Image 2 was released.
The entire AI design circle just exploded.
How powerful is the effect? Let's put it this way: the phrase "seeing is believing" might become a complete joke from now on.
I couldn't help but post on social media to give everyone a friendly reminder:
It's no exaggeration to say that half of the designers in my circle have already started trembling.
Regarding the article about DeepSeek V4 yesterday, many people asked how the cover image was made.
In fact, it was generated directly by GPT-Image 2 in one click, without any post-processing.
I spent some time collecting interesting gameplay cases and prompts from the web, extracted them into templates, and open-sourced them on GitHub.
There are so many categories, covering mainstream types like infographics, interfaces, posters, photography, and illustration, as well as supplementary cases like branding, cards, livestream screenshots, Chinese style, and commercial visuals.
There are many cases under each category, and of course, every case comes standard with prompts:
Most importantly, we have extracted industrial-grade prompt templates.
Each set of templates comes with a "pitfall prevention guide," which you can directly give to your Agent to learn, allowing for batch prompt generation and batch image production in one go.
Recommended to be called directly by Agents like Claude Code or Codex.
The open-source repository address is here. If you find it helpful, please give it a star, haha.
1https://github.com/freestylefly/awesome-gpt-image-2
In addition to those collected online, I have also tested many cases myself and attached the prompts.
Let everyone feel just how powerful GPT-Image 2 really is.
1. Livestreaming Screens
This kind of image is incredibly realistic.
Here's another one:
It's truly beautiful.
The prompt is very simple:
Generate an image of a livestreaming room. The atmosphere is a beautiful woman dancing under the moon, and there are many people commenting in the room.
To cater to female readers, here's a handsome guy selling Hanfu.
Prompt:
Generate an image of a livestreaming room. The atmosphere is a handsome guy explaining Hanfu, and there are many people commenting in the room.
2. Hand-drawn Style City Maps
Xi'an:
Previously, when using other AI tools, to be honest, they couldn't stand close scrutiny, and the text garbling issue was very serious.
GPT-Image 2 is much better regarding Chinese text garbling.
Here's one for Xinjiang:
Template prompt is super simple:
Generate a hand-drawn watercolor style city map of [Insert City Name], including local specialty foods, landmarks, and city characteristics.
3. Posters
I just took a quick photo of a bottle of Tea π and gave it to GPT-Image 2.
The resulting poster was, honestly, much better than I expected.
Prompt:
Help generate a promotional image for this product.
With just that one sentence, it imagined all the design details on its own.
4. Exploded Views
Generate an exploded view of AI glasses:
Prompt:
Generate an exploded view of AI glasses, including the names of each component and several core selling points of this product.
5. Technical Explanation Diagrams
For example, this RAG technical explanation diagram:
No matter what technology you want to explain, you can use this template:
Help me generate a detailed explanation diagram for [Insert Term Here] technology.
6. Social Media Moments
It can generate very realistic screenshots of social media moments.
So, I strongly suggest you forward this article to your parents; the cost of fake photos is really too low now.
Creating an X post takes just a minute:
I won't post the prompts; honestly, it's quite scary when you think about it.
7. Personal Webpages
I asked it to design a personal homepage for me, and the result was quite good:
8. Poetry Imagery
For example, this "Duan Ge Xing," hahaha:
Prompt:
Help me generate an artistic image for "Duan Ge Xing," including the full text of the poem.
9. Long Scroll Images
GPT-Image 2 supports many direct output sizes, and long scrolls are no problem at all.
Moreover, there is almost zero text garbling. It's too strong.
Prompt: Help me generate a long scroll image for "Thinking of the Red Cliff," including the full text of "Ode to the Red Cliff."
GPT-Image 2 is now available in many places, such as Lovart, which updated immediately. You can go try it.
If you have a GPT Plus membership, you can use it directly on the web.
There's an even cooler way: you can have Codex call GPT-Image 2 directly to generate images.
Previously, you had to coordinate with some skills, but now that's not necessary at all.
And I found that Codex + Obsidian is the best partner for official account cover images.
Codex directly calls GPT-Image 2 to generate images. After finishing the article in Obsidian, the cover image is generated conveniently and can even be automatically filled into the article fields.
GPT-Image 2's understanding is truly strong.
Previously, using Claude Code + Nano Banana 2 to generate images was slow and unstable.
Now, with Codex generating a cover image, it takes less than a minute. It's so comfortable.
This is how my current creative process and cover image generation work.
Alright, after looking at the cases, let's talk about why I made this open-source project.
Recently, after GPT-Image 2 went into full release, the AI drawing trend has been huge.
I accidentally stumbled upon the prompt libraries of YouMind and OpenNana. I didn't pay much attention at first.
But when I clicked in, wow, the pictures were really beautiful.
However, after messing around for a while, I found a very awkward problem.
Currently, this kind of simple prompt aggregation and collection is, to put it bluntly, just displaying a bunch of exquisite specimens.
If I want to integrate hundreds of prose-like prompts into my own Agent or automated workflow, do I still have to manually scrape them one by one?
Can't we let AI do this tedious work?
I don't want a bunch of sentences; I want a set of "protocols" that can be called directly.
So I spent some time reverse-engineering all 329 exquisite cases.
To be honest, at first, I just wanted to make a "deepened Chinese version."
But after it was up and running, the results were incredibly good, especially when used with Agents.
I realized we have completely changed the way of playing.
This is no longer a simple translation of someone else's library; it's an industrial-grade Prompt-as-Code engine designed specifically for GPT-Image 2.
Briefly, here's what I did:
🧱 Atomic Schema Injection: Stop Writing Essays
I reduced all visual elements (subject, lighting, material, layout) into JSON/YAML structured components.
Your Agent can parse them stably with its eyes closed, with zero hallucinations.
⚙️ Zero-Configuration Workflow
Lowering the barrier to entry, seamlessly connecting to your large model data pipeline at any time.
Honestly, watching them coordinate to work and produce images gives a real sense of satisfaction as a client.
🧬 Multi-dimensional Decision Matrix
Targeting GPT-Image 2's powerful text layout capabilities, I introduced precise spatial coordinate system constraints.
Completely eliminating the technical blind spot where traditional NLP couldn't control screen layout.
It's easy to fall into a rut in technology: if others make a gallery, I make a gallery.
But there are always moments when you have to stop and ask yourself: Can this be automated? Can it become a foundation?
Aligning with the original version later might just be a matter of adding a few automation scripts.
But for now, we are trying to develop a methodology for prompt templating.
If you are also tired of the "gacha" style image generation experience, come and try it.
Final Words
To be honest, I was quite nervous when I started this open-source project.
After all, there are already enough AI drawing libraries out there. Does everyone really need another one?
But when I ran the first image using this set of templates myself, watching the Agent generate images fully automatically, with zero hallucinations and in one shot, I suddenly felt it was worth it.
The world never lacks good prompts; what it lacks is a methodology to turn good prompts into something "available to everyone and callable by Agents."
If you are also a fellow traveler tinkering with AI workflows, feel free to give me a Star on GitHub, and you're welcome to submit your own good cases. Let's make this library better and better together.
See you next time. I hope we can all do something different with AI.
