Open-Source 5B Model Beats GPT Image 2.5 in Complex Chinese Calligraphy Test

Today, DeepSeek open-sourced its production-grade inference framework, DeepInfer. Many are discussing how to run stronger models with less compute.

Coincidentally, I've been testing another open-source model following the same path: MOSS-Image, which has only 5B parameters. I gave it and the supposedly strongest closed-source model, GPT Image 2.5, the same difficult challenge.

The result: GPT Image 2.5 failed on the second round, while the open-source 5B small model correctly rendered the entire poem "Bring in the Wine" twice. If Li Bai saw this, he'd probably drink three hundred more cups.

The prompt was: Create a scene of Li Bai raising a cup by the river in an ink-wash scroll painting. Write the full text of "Bring in the Wine" in 12 lines of calligraphy, accompanied by a title, author name, one vermilion seal, and three lines of English text. There must be exactly 4 wine cups on the stone platform—no more, no less.

I originally thought the closed-source model would win easily. In the first round, both were comparable. But in the second round, when I changed just two details, GPT Image 2.5 miscounted the cups, while MOSS-Image got it right both times.

Below are 2 images generated by MOSS-Image

Below are 2 images generated by GPT Image 2.5

Why is this test hard?

I intentionally wrote this prompt to be difficult.

This famous poem has 12 lines and over 170 characters, which need to be written in running script on the painting without a single error. I've tested many models before; once the character count gets high, they start making mistakes.

The number of wine cups was added specifically because AI image generation is notoriously bad at counting objects.

Plus, the scene includes many elements: river landscape, Li Bai, title, seal, the whole poem, and three lines of English, each with specific positions. I also added a constraint at the end: no other text allowed, and only one seal.

Round 1: Both Passed

In the first round, I asked for 2 wine cups on the stone platform, and the last line of English was "Let us drink away the sorrows of all ages."

When the images came out, I checked them one by one: no errors in the poem, only one seal, correct English text, and exactly 2 cups. Both models passed this round smoothly.

The difference was mainly in aesthetics. The MOSS-Image version had varied ink density and comfortable white space, looking like a real hanging scroll. GPT Image 2.5 looked good at first glance, but zooming in revealed flatter brushstrokes, lacking the subtle bleeding effect of traditional ink wash.

Round 2: Small Changes Revealed Big Differences

Next, I changed only two things:

1. Changed the number of cups from 2 to 4.
2. Added the word "Cheers" after the last line of English.

Logically, these are minimal changes. The results were:

• MOSS-Image: Got both changes right. Exactly 4 cups, "Cheers" added, and the 12 lines of poetry and the seal remained intact.
• GPT Image 2.5: Added "Cheers," but messed up the cup count. It required 4 cups but drew 5.

Of course, this is just one set of prompts and doesn't represent all scenarios. But regarding "long text plus precise control," the open-source 5B model delivered an impressive performance.

How Can a 5B Small Model Do This?

Looking back at this test, MOSS-Image's success in both rounds relies on several capabilities:

• Accurate Chinese/English Rendering. The 12 lines of over 170 Chinese characters in running script had no typos or omissions. The bottom three lines of English and the added "Cheers" were also clear.
• Complete Understanding of Long Prompts. It supports prompts up to 4K tokens. My prompt included style, scene, title, poem, English, and negative constraints. It followed everything from start to finish without losing track. So when I changed the cup count and added a word in round two, it handled both updates without disrupting the rest.
• Controllable Style. This test required a rice-paper ink-wash texture, and it didn't drift into photorealistic or 3D styles. Besides traditional Chinese style, it can also handle anime, illustration, and photo-realism.

If you want to use it in your work, here are two points worth knowing: First, it supports structured JSON Layout, allowing direct specification of text and element positions, which improves composition control for posters, e-commerce images, and social media covers. Second, weights, code, and usage guides for both Base and SFT versions are open-sourced. SGLang and vLLM-Omni provide Day 0 support, making LoRA fine-tuning or secondary development easy to start.

Testing Method and Prompt Tips

The process was simple: both models used the same prompt. After running round one, I changed two details and ran round two. MOSS-Image was generated on the official experience platform, while GPT Image 2.5 was generated online in ChatGPT.

For writing complex prompts like this, I have three tips:

1. Define Art Style and Material First. Specify rice paper, ink wash, mineral pigments, and add "not photographic, not 3D" to prevent the image from drifting into realism.
2. Describe by Regions. Describe the scene in the lower left, title and seal in the upper right, poem in the right half, and English at the bottom. Separating blocks makes it easier for the model to follow.
3. End with Negative Constraints. Add "No other text, no extra seals, no watermarks" at the end to keep the image clean.

Full Prompt

Here is the complete prompt used in the first round, ready to copy and try:

1A traditional Chinese ink-and-wash painting on warm ivory xuan paper, tall hanging-scroll format. Layered black ink, soft gray washes, muted blue-green mineral pigment, pale ochre, visible paper fibers, feathered wet-ink edges, dry-brush texture. Hand-painted, not photographic, not 3D.
2Scene in the left and lower part: a mighty river bursts from cloud-veiled mountains, drawn with flowing ink lines and blank paper. Weathered cliffs, layered distant peaks, a wind-bent pine, drifting mist, a pale moon behind thin clouds. On a rocky riverside terrace, a poet in loose ivory robes with a simple topknot raises a small wine cup to the sky. A round ceramic wine jar and two cups sit on a flat stone beside him.
3Upper right: large bold brush title 《将进酒》, below it the author name 李白, with exactly one small vermilion seal reading 诗酒 next to the author name.
4Right half: the complete poem in clean, clearly readable regular-running script Chinese calligraphy, written in exactly 12 horizontal lines, even line spacing, every character correct, on clear blank paper:
5Line 1: 君不见，黄河之水天上来，奔流到海不复回。
6Line 2: 君不见，高堂明镜悲白发，朝如青丝暮成雪。
7Line 3: 人生得意须尽欢，莫使金樽空对月。
8Line 4: 天生我材必有用，千金散尽还复来。
9Line 5: 烹羊宰牛且为乐，会须一饮三百杯。
10Line 6: 岑夫子，丹丘生，将进酒，杯莫停。
11Line 7: 与君歌一曲，请君为我倾耳听。
12Line 8: 钟鼓馔玉不足贵，但愿长醉不复醒。
13Line 9: 古来圣贤皆寂寞，惟有饮者留其名。
14Line 10: 陈王昔时宴平乐，斗酒十千恣欢谑。
15Line 11: 主人何为言少钱，径须沽取对君酌。
16Line 12: 五花马，千金裘，呼儿将出换美酒，与尔同销万古愁。
17Bottom center, small dark-gray serif English text in three lines:
18BRING IN THE WINE
19Li Bai
20Let us drink away the sorrows of all ages.
21No other text, no extra seals, no watermarks anywhere else in the painting.

For the second round, only two changes were made:

1. Changed "two cups" to "four cups" in the scene description.
2. Added "Cheers" after the last line of English.

Everything else remained unchanged.

Your Turn

An open-source 5B model can write the entire poem "Bring in the Wine" into an ink-wash scroll, getting the cup count exactly right.

Currently, MOSS-Image has surpassed 700,000 downloads on ComfyUI. Don't just take my word for it, try it yourself 👇

🤗 Free Online Experience: https://huggingface.co/spaces/OpenMOSS-Team/MOSS-Image

⭐ GitHub for Weights and Code: https://github.com/OpenMOSS/MOSS-Image

What will you create with it? Share your images in the comments. Let's raise our cups together. Cheers 🍶

#MOSS-Image #AIGC