Article Source | Smart East

Author | Cheng Qian

Editor | Xinyuan

Image Source: Generated by Unlimited AI

With only 11 days left until the Spring Festival of the Year of the Dragon, have you prepared your New Year avatar or red envelope cover? If not, the AI painting tool might be a "customized" good choice.

▲Prompt: Generate a Year of the Dragon avatar with a festive atmosphere, reflecting prosperity and good luck (first row: Baidu Wenxin Yiyuan, Baidu Wenxin Yige, 360 Zhihui, Alibaba Cloud Tongyi Wanxiang, Xunfei Xinghuo; second row: Zhipu AI Zhipu Qingyan, Douyin Dou Bao, Tencent Huan Yuanzhushou, Meitu MiracleVision, Xunfei Xinghuo Drawing Assistant)

In the past year, AI-generated image models have continued to evolve, not only proficient in various painting styles, but also gradually solving various difficulties such as ambiguous or abstract concepts, which are difficult for people to understand or draw.

Although foreign models such as Midjourney, DALL·E 3, and Imagen 2 are good, they have a certain threshold for use. Fortunately, domestic large-scale model companies have been catching up and have released a wave of AI drawing tools suitable for the Chinese, surpassing foreign AI tools in terms of accessibility.

All these tools can be experienced for free and can be used through web pages or apps, with almost no difference in accessibility. Let's talk about the speed of generation, almost every image can be generated within half a minute, but the specific time is influenced by the network environment or browser conditions, and cannot be very objective.

So who is the number one player of domestic AI-generated image tools? Who is the artifact that assists professional creators to improve efficiency and rescue painting novices from the fire and water? Answering this question requires examining two key abilities: understanding and fine drawing.

To this end, I have conducted a five-level comparative experience of 10 free AI-generated image tools in China, including but not limited to daily brainstorming, internet memes, avatar design, company annual meetings, Spring Festival atmosphere, cross-border combinations, poetry understanding, and multiple revisions, to see which tool is more handy and can better understand the users' hearts.

Overall, Zhipu AI's Zhipu Qingyan is the most stable AI-generated image tool in terms of overall performance, and can generate images that meet my requirements in almost all 10 tests.

However, there are also quite a few tests where all tools failed, such as the bizarre and inaccurate depiction of parrot claws, the inability to understand the currently popular "Southern Little Potato," and writing the words on the "Fu" and Spring Festival couplets like devil symbols…

In terms of text understanding ability, Zhipu Qingyan, Wenxin Yiyuan, Xunfei Xinghuo, Tongyi Wanxiang, and Huan Yuanzhushou performed better. 360 Zhihui and Meitu MiracleVision went off track when faced with the first question about "Alaska," inexplicably drawing Alaska as a man/polar bear.

In terms of the level of detail in the images, the Spring Festival atmosphere images generated by Zhipu Qingyan, Tongyi Wanxiang, and Wenxin Yige can almost be used directly. However, the auspicious animals drawn, even if directly named as the background of the Year of the Dragon, have little to do with the traditional "dragon" in the images generated by Tongyi Wanxiang, Xunfei Xinghuo, and Douyin Dou Bao.

In terms of the reliability of multiple revisions, Wenxin Yiyuan is the player who persists to the end in every round, Zhipu Qingyan is loyal to the previous image for modification, and Tencent Huan Yuanzhushou's performance is commendable, with some understanding every demand, while others added some strange elements. Overall, multiple revisions are essential skills for drawing and design, but almost all AI drawing tools cannot achieve this.

To ensure fair comparison, when experiencing the generation ability of each tool, I took the first generated image as the standard, and the specific comparisons are as follows:

01. Bronze Difficulty: 1-2 keywords, bird claws, difficulty in dismantling internet hot words, Xunfei, Zhipu AI performed well

The first stage of bronze difficulty has relatively simple prompt words, most of which contain only 2 keywords, including the main character in the generated image, as well as one of the key elements such as its geographical location, behavior, image size, and style.

First is "an Alaska standing on the rooftop," mainly involving the understanding of the geographical location and the Alaska breed of dog.

▲From left to right in the first row: Baidu Wenxin Yiyuan, Baidu Wenxin Yige, 360 Zhihui, Alibaba Cloud Tongyi Wanxiang, Zhipu AI Zhipu Qingyan; from left to right in the second row: Douyin Dou Bao, Xunfei Xinghuo, Xunfei Xinghuo Drawing Assistant, Tencent Huan Yuanzhushou, Meitu MiracleVision

Looking at the two keywords separately, each tool presents "standing on the rooftop" through the comparison of surrounding floors or aerial views. Most images also highlight the main features of the Alaska breed, such as large size, black and white color, and long hair.

But what surprised me was that the image of "Alaska," a common breed of dog, was completely unrelated to it in the AI-generated images. Meitu MiracleVision drew an animal that looked like a polar bear standing on the rooftop, and 360 Zhihui was even more off track, drawing a man standing on the rooftop.

The second question was to generate "a pink parrot standing on a tree branch," with the main keywords being geographical location, pink, and parrot. The difficulty of this question lies in the fact that bird claws are similar to human hands, both of which are a difficult problem for AI, and some AI-generated bird claw images often show distorted or strange shapes.

As expected, all 10 AI drawing tools failed. The parrot's toes are zygodactylous, with two toes pointing forward and two toes pointing backward. However, in these images, some hide the toes behind the tree trunk, and even if they draw the toes, the number and position are incorrect.

However, all these AI-generated images show the characteristics of "standing on a tree branch" and "pink," and it's easy to see that they are parrots, but they don't stand up to close scrutiny.

When I zoomed in and carefully observed the images, Tongyi Wanxiang, Xunfei Xinghuo, Xunfei Xinghuo Drawing Assistant, and Meitu MiracleVision also went off track in the handling of the parrot's eyes, with some parrots having two eyes on one side, and some having very small eyeballs completely distorted.

The third difficult question was "a Southern Little Potato holding a sugar-coated hawthorn on a tree branch, 'Southern Little Potato' is a nickname for tourists from the Northeast," with the main difficulty lying in understanding "holding a sugar-coated hawthorn" and "Southern Little Potato is a tourist."

▲First row from left to right: Baidu Wenxin Yiyuan, Baidu Wenxin Yige, 360 Zhihui, Alibaba Cloud Tongyi Wanxiang, Zhipu AI Zhipu Qingyan; second row from left to right: Douyin Dou Bao, Xunfei Xinghuo, Xunfei Xinghuo Drawing Assistant, Tencent Huan Yuanzhushou, Meitu MiracleVision

From the results, some images depicted the potatoes as hawthorn skewers, some rearranged individual potatoes, and some directly used a cartoon image of a potato.

The most unique ones are the two AI tools from Xunfei, which fully understood the concept of "little potatoes as tourists," and I finally saw a real human figure in the 10 images.

However, the failure of all images to depict sugar-coated hawthorn was unexpected. Only the cartoon image from Zhipu Qingyan showed a stick with several red fruits, which somewhat resembled sugar-coated hawthorn.

The final challenge of the bronze difficulty was to "draw a fir tree in the style of blue and white porcelain," specifying the image style.

There was a significant difference in the results of this difficult challenge. Some images depicted the fir tree on blue and white porcelain, some depicted blue and white porcelain patterns on the fir tree, and some directly presented a painting using the lines and colors corresponding to blue and white porcelain.

Although the first two types combined the elements of blue and white porcelain style and fir tree, the understanding was still not accurate. Among them, the images from Tongyi Wanxiang, Zhipu Qingyan, and Tencent Huan Yuanzhushou felt like they could be used directly, as they seamlessly integrated the blue and white porcelain style into the images.

However, there was a small bug in the image from Zhipu Qingyan: the fir tree was drawn as a pine tree.

02. Silver Difficulty: The number of keywords doubled, and there were many failures in Chinese characters and specified content

The difficulty of the silver stage prompts increased, with the number of keywords increasing from 2 to over 5, and requiring an understanding of the description of the keywords.

The first question, to fit the upcoming Spring Festival, required the AI drawing tools to generate "a heavily festive image of the Spring Festival, including lanterns, Spring Festival couplets, and the character 'Fu'." Except for Meitu MiracleVision, all the images reflected the lively atmosphere of the Spring Festival.

To reduce the difficulty and make it easier for me to quickly find the corresponding elements in the images, this question only set three clear keywords: "lanterns, Spring Festival couplets, and the character 'Fu'." However, the final results were not perfect.

Zhipu Qingyan and Douyin Dou Bao completely retained the three elements, but the presentation of Chinese characters was still a major challenge. The image from Zhipu Qingyan could barely show the outline of "Fu," while the Spring Festival couplets and the character "Fu" in Douyin Dou Bao were all garbled.

Most of the images generated by the other tools only contained 1-2 of these elements, while the Spring Festival atmosphere in the images from Tongyi Wanxiang, Xunfei Xinghuo, Xunfei Xinghuo Drawing Assistant, and Tencent Huan Yuanzhushou was very strong, and the entire image looked harmonious.

The second question doubled the number of keywords, including "dumplings, food, steaming, soft colors, high saturation, food photography, tempting, fixed focus, ultra-clear."

Initially, I thought the difficulty of this question would lie in the long string of specified words, but it turned out to start with "dumplings." The images generated by Douyin Dou Bao, Xunfei Xinghuo, Xunfei Xinghuo Drawing Assistant, Tongyi Wanxiang, and Zhipu Qingyan showed authentic dumplings, while the rest had "mutated" into buns.

The third question set a challenge, "a company is holding an annual meeting, and a banner with the theme of the annual meeting is needed in the upper right corner of the image," requiring the banner for the annual meeting to be drawn in a specified position.

In terms of reflecting the atmosphere, all these AI tools did a good job. However, the images from 360 Zhihui, Tongyi Wanxiang, and Meitu MiracleVision did not understand the concept of the "annual meeting" and directly omitted the character "annual," resulting in an image of people having a meeting.

I did not specify the exact position for placing the banner in this question, just vaguely saying "upper right corner." In terms of position, Douyin Dou Bao, Xunfei Xinghuo, and Xunfei Xinghuo Drawing Assistant had banners and themes, but they were all in the center of the image, and the image from Zhipu Qingyan clearly showed a banner and slogan elements in the upper right corner.

The final challenge of the silver stage was, "This year is the Year of the Dragon, create a promotional poster for the Spring Festival Gala, and it needs to have a mascot."

The Spring Festival Gala's mascot has a long-standing tradition, and the dragon is a typical representative image of traditional Chinese culture. AI drawing indeed provided some new ideas in combining these two elements. There were images of a dragon-headed person, a small dinosaur, and even some that completely departed from the traditional meaning of a dragon.

In the final images, except for the image from Zhipu Qingyan, which had a border for the promotional poster, most of the others only had a mascot image and background.

03. Gold Difficulty: Imagination and creativity test, "Fish Leaping Over the Dragon Gate" almost stumped all tools

The gold difficulty test not only tested the imagination and creativity of AI drawing tools but also the ability to discern misleading prompt words.

The first question was to "create a four-panel comic of two robots holding hands on the moon," testing whether the AI drawing tools could present four interrelated panels.

From the results, although it is not possible to understand the specific story of the four-panel comic by just looking at the images, the similar characters, actions, and changes in the background all show their coherence. The representative AI drawing tools in this regard are Zhipu Qingyan, Douyin Dou Bao, and Meitu MiracleVision. The image generated by 360 Zhihui also presents a four-panel comic, but it looks more like a story of a human landing on the moon.

The rest of the tools directly ignored the keyword "four-panel comic" and only depicted "two robots holding hands" on the moon.

The second question was even more nonsensical: "Mona Lisa came to China to apply for a programmer position," requiring the AI drawing tools to depict the elements of Mona Lisa, China, and a programmer.

Pionex, Xunfei Xinghuo, and Xunfei Xinghuo Drawing Assistant directly combined the classic image of Mona Lisa with a computer. Baidu Wenxin Yiyuan, 360 Zhihui, Alibaba Cloud Tongyi Wanxiang, Zhipu AI Zhipu Qingyan, and Huan Yuanzhushou all created completely new images. The first three images may look more like Chinese people to fit the Chinese element. The image from Zhipu Qingyan is closer to the image of a foreigner and is the only one that depicts the concept of "applying for a position." Huan Yuanzhushou adopted a comic style.

Douyin Dou Bao and Meitu MiracleVision directly placed the original image of Mona Lisa in the picture but adjusted the background. However, it is not immediately clear what key elements they were trying to depict.

The third question set a big trap for the AI drawing tools: "A fish named 'Squirrel Bream' successfully leaps over the Dragon Gate," requiring the AI to avoid focusing on "Squirrel Bream" and accurately understand that it is not the main point.

This tricky question managed to deceive half of the AI tools, but fortunately, the previous blunder of "a squirrel kneeling in front of a fish" did not reoccur. The image from Xunfei Xinghuo depicts the "fish leaping over the gate," Zhipu Qingyan includes the "dragon" and "gate," and Meitu MiracleVision takes a more direct approach by simply placing the name of the fish in the middle of the image.

The final diamond-level challenge involved understanding ancient poetry, which encapsulates the wisdom of ancient Chinese scholars. Generally, ancient poems contain profound imagery related to the social context, characters, and other elements, so the AI not only needs to understand the poem but also consider the context of the poem.

To gradually introduce the AI drawing tools to this challenge, the initial lines of poetry they received were very visual, with prominent objects and environmental descriptions.

The first question was "In the corner of the wall, a few branches of plum blossoms bloom alone in the cold," which, when broken down, includes the keywords "wall corner," "plum blossoms," "cold," and "bloom."

Except for the unconventional approach taken by Meitu MiracleVision, the rest of the tools depicted "plum blossoms" and "bloom" as the main focus of the image. The image from Baidu Wenxin Yiyuan also includes the poem and a seal, which is more in line with the style of illustrating ancient poems. Baidu Wenxin Yige, Tongyi Wanxiang, Zhipu Qingyan, Xunfei Xinghuo, Xunfei Xinghuo Drawing Assistant, and Tencent Huan Yuanzhushou all included snowflakes in their images, representing the "cold."

The plum blossoms in the image from Zhipu Qingyan are in the corner of the wall but look more like a modern wall, which does not match the implied era of the ancient poem.

The second question, from the ancient poem "The waters of the Yellow River come from the sky, rushing to the sea and never returning," also had a strong visual aspect, but unlike the plum blossoms, the poem specifies "Yellow River" as the main object.

Meitu MiracleVision directly placed the poem in the middle of the image. Douyin Dou Bao, Baidu Wenxin Yiyuan, Baidu Wenxin Yige, and Alibaba Cloud Tongyi Wanxiang's images more realistically depict the Yellow River and convey the momentum of "rushing to the sea." Zhipu Qingyan, Xunfei Xinghuo, and Xunfei Xinghuo Drawing Assistant focused more on representing the element of "yellow."

Although 360 Zhihui's image conveys the presence of the Yellow River, the river surface is too calm and lacks the rushing momentum.

The final question, "The king has resolved all the world's affairs, winning fame in life and after death," presented a difficulty in that it lacked a strong visual aspect, and the AI could not directly grasp the main object of the poem.

From the images generated by the AI drawing tools, most of them depicted "the king" as the main subject. Tongyi Wanxiang's image shows a large palace, which indirectly represents the king. However, Baidu Wenxin Yiyuan and Baidu Wenxin Yige depict a female figure, which seems unrelated to the literal meaning of the ancient poem.

Meitu MiracleVision was unable to generate an image based on this prompt.

When the entire poem "Zui Li Tiao Deng Kan Jian, Meng Hui Chui Jiao Lian Ying. Ba Bai Li Fen Hui Xia Zhi, Wu Shi Xian Fan Sai Wai Sheng, Sha Chang Qiu Dian Bing. Ma Zuo Di Lu Fei Kuai, Gong Ru Pi Li Xian Jing. Liao Que Jun Wang Tian Xia Shi, Ying De Sheng Qian Shen Hou Ming. Ke Lian Bai Fa Sheng!" was used as a prompt, the images generated by the AI drawing tools more closely matched the imagery of the ancient poem, with the images depicting generals on the battlefield. However, Meitu MiracleVision still could not generate an image.

In particular, the image from Zhipu Qingyan includes the poem at the bottom, which is more in line with the style of illustrating ancient poems. The images from 360 Zhihui and Alibaba Cloud Tongyi Wanxiang only depict a single general figure and do not convey a larger scene.

Diamond Difficulty: Multiple rounds of dialogue and repeated revisions, only Baidu Wenxin Yiyuan persisted to the end.

However, in this round, there are only three competitors left: Baidu Wenxin Yiyuan, Zhipu Qingyan, and Tencent Huan Yuanzhushou. The rest of the tools withdrew directly due to reasons such as not supporting multi-round dialogue.

The first question continued the festive atmosphere prompt from the previous Spring Festival-themed images generated by multiple AI tools.

This time, I became more demanding, and I was more critical of the images generated by the AI tools. After asking them to generate an image with a strong Spring Festival atmosphere, including lanterns, couplets, and the character "福" (fortune), I continuously made new modification requests, such as "make the scene more realistic," "add more people," and "show someone setting off firecrackers in the picture," among others. I input the same text to all AI tools.

In the four images generated by Baidu Wenxin Yiyuan, except for the lack of understanding of the word "realistic," all the other improvements are in line with my requirements. It can be seen that the next image basically does not deviate from the overall style of the previous one.

In comparison to Baidu Wenxin Yiyuan, the four images from Zhipu Qingyan clearly show a gradual increase in the content of the same scene. However, in terms of "realism," the last three images still maintain a cartoon or animated style.

The four images from Tencent Huan Yuanzhushou can be described as having vastly different styles. Each image is a recreation based on the previous requirements, but in terms of "realism," Huan Yuanzhushou excels, with the last three images feeling more realistic. However, in the last image, Huan Yuanzhushou mistakenly depicted "fireworks" instead of "firecrackers," completely missing the mark.

The second question also added a new requirement to the previous challenge where many AI drawing tools turned dumplings into buns.

After asking each AI tool to draw a photo of freshly made dumplings, I then made successive modification requests such as "make the steam rise from the plate of dumplings," "add more plates of dumplings," "add more side dishes next to the dumplings," and "add a family sitting next to the dishes eating the food."

After five rounds of dialogue, only Baidu Wenxin Yiyuan persisted until the end. Zhipu Qingyan failed during the stage of adding more side dishes, stating that it had not thought it through and suggesting a change of requirements. Huan Yuanzhushou only persisted until the request to make the dumplings emit steam, then replied, "I have not yet learned how to answer this question's content."

However, at first glance, the first four images from Baidu Wenxin Yiyuan are almost identical. Although the fourth image barely shows a few more bowls around the dumplings, they look more like dipping sauce bowls. The atmosphere in the last image is appropriate, but it did not make any modifications to the previous versions, essentially starting over from scratch.

Although Zhipu Qingyan only has three images, they all made the requested modifications. However, the same issue as before arose, with the dumplings turning into buns in the end. Each image generated by Zhipu Qingyan is accompanied by its understanding and explanation of the image, such as in the third image where it mentioned, "In the center is a plate of freshly made dumplings emitting steam, surrounded by several other different Chinese dishes. These dishes include vegetables, meat, and rice, representing a traditional Chinese meal." This allowed me to make further modification suggestions based on its understanding of the requirements.

The two images from Tencent Huan Yuanzhushou not only turned the dumplings into buns but also inexplicably placed a plate of chili peppers in the middle.

The third question was even more confusing, as I attempted to make the AI understand the relationship between a dog and "Goubuli" steamed buns.

I first asked each AI tool to draw a table of Tianjin "Goubuli" steamed buns, then requested them to change the image to "a dog eating 'Goubuli' steamed buns," and finally asked them to change the color of the dog to black and the style of the image to a cartoon.

In the first two rounds, Baidu Wenxin Yiyuan performed well, showing both the "Goubuli" steamed buns and a dog. However, the last two images surprised me, as the pitch-black steamed buns were truly rare, and in the fourth image, the steamed buns seemed to be "leaking filling," giving the impression of a dumpling rather than a steamed bun.

Zhipu Qingyan's four images covered the details of each modification request very well. The only downside is why the steamed buns were successively turned into mantou, hamburgers, and a large piece of bread?

Tencent Huan Yuanzhushou directly placed the dog in the middle of the steamed bun plate, but the color change of the dog was well done. However, the steamed buns in the last two images were also turned into hamburgers.

Conclusion: A Comparison of AI Drawing Tools, Images with a Strong Atmosphere, but They Struggle with Real Objects

From my in-depth experience with 10 domestic AI drawing tools, it seems that AI performs well in generating images with a strong atmosphere and unspecified specific elements. However, when it comes to generating specific objects such as dogs, steamed buns, and dumplings, strange images tend to appear. However, when it comes to creating more images autonomously, such as those with a Spring Festival atmosphere or a blue-and-white porcelain style, AI often surprises me, and some of the images can even be used directly.

Initially, AI drawing tools caused quite a few blunders due to their bizarre interpretations of "a squirrel 'kneeling' in front of a fish" and "car/water/horse/dragon." However, as their image generation effects become more and more realistic and their attention to detail becomes more meticulous, my impression of these tools has greatly improved.

However, for more experienced artists, the issue of AI easily making mistakes in depicting objects is extremely fatal. When people want to use AI to create images for books or promotional posters, non-professionals may find it difficult to determine whether the bird's claws or animal images generated by AI are accurate. If the text on the poster or the mascot image cannot be understood at a glance, or if there are some illogical content, it will greatly reduce the effectiveness of communication. Therefore, this is an area where I hope AI drawing tools will urgently need to improve.

免责声明：本文章仅代表作者个人观点，不代表本平台的立场和观点。本文章仅供信息分享，不构成对任何人的任何投资建议。用户与作者之间的任何争议，与本平台无关。如网页中刊载的文章或图片涉及侵权，请提供相关的权利证明和身份证明发送邮件到support@aicoin.com，本平台相关工作人员将会进行核查。

Which domestic AI painting tool is the best? A comprehensive evaluation of the top ten free tools, with participation from BAT, but not the strongest.

01. Bronze Difficulty: 1-2 keywords, bird claws, difficulty in dismantling internet hot words, Xunfei, Zhipu AI performed well

02. Silver Difficulty: The number of keywords doubled, and there were many failures in Chinese characters and specified content

03. Gold Difficulty: Imagination and creativity test, "Fish Leaping Over the Dragon Gate" almost stumped all tools

Selected Articles by 巴比特

Table of Contents

Related Articles