Yun Zhong sent from Aofei Temple
Qubit | Public Account QbitAI
The first batch of designers who have been robbed by AI have already appeared.
As various major Internet companies have announced how to integrate AI tools into their daily workflow, there have been layoffs in many positions such as art, design, and original painting. There is no doubt that the cost reduction and efficiency increase of this wave of AI has really begun to affect everyone. rice bowl.
Just recently, an overseas generative marketing software called ZMO.AI has rapidly surpassed one million monthly B-end users , and its ARR has reached 3 million US dollars.
△ Official website address:https://background.zmo.ai/
It seems that old marketers can't escape this wave of AI unemployment.
It is reported that the AI background generation under ZMO.AI only needs the merchant to upload a product picture, and it can generate thousands of different styles of backgrounds according to the instructions on the premise of retaining 100% product details .
Its fidelity is comparable to that of a large-scale commercial scene map, whether it is light and shadow or clarity, it is better than a PS master with more than 10 years of experience.
Another product of the company, Marketing Copilot , only needs to upload a product picture, from shooting, to poster production, to post-production optimization, all embedded in the automated process of AI workflow, using AI's powerful creativity and analysis capabilities to achieve Operation is optimized every second.
Such an out-of-the-box product allows small bosses who don't know how to shoot and PS to get started.
△ Official website address:https://background.zmo.ai/
In fact, ZMO.AI has been rooted in the field of marketing generation for a long time.
Since the end of 2020, the AI model function has been launched in the field of clothing marketing, and there have been many in-depth and successful cooperations with well-known domestic clothing brands.
With the fire of Diffusion, ZMO also took advantage of the trend to launch the AI content generation product ImgCreator.AI for a wider audience in September 2022, and took the marketing crowd as the main service object, strengthening the background generation, poster generation and data optimization AI in the product ability, and provide Marketing Copliot value-added functions for B-end users.
This move has won ZMO more than one million monthly active high-value small B-end users, and the users have a strong willingness to pay, and there is good news that the ARR will quickly reach 3 million US dollars.
The generated product is wrong?
Different from other C-end users of AI painting purely for fun, B-end users are faced with very professional scenes. Compared with C-end users, the requirements for quality, controllability and accuracy are extremely high. High, this may also be the reason for the success of specialized AI content products like ZMO.
Rowdy is the CEO of e-Bike, a British start-up company. They are a small team of less than 10 people. Their product e-bike focuses on the anti-theft system for electric bicycles.
According to Rowdy, for small companies, the large amount of material required for website construction and blog writing is very expensive, and the emergence of AIGC has greatly liberated their productivity.
However, Rowdy found that a large number of AIGC websites are often in an artistic aesthetic style, which is far from the real photo style he needs, while the real photo style of ZMO.AI is very realistic, and the resolution can reach 4/8K, which is completely invisible It's an image generated by AI.
In the past six months, Rowdy's team has been using ZMO's products to provide pictures for website design and company blog, and can generate more than 200 pictures every week.
According to Rowdy's description:
Compared with the expensive shooting, the software fee of more than 20 pounds is simply too cost-effective.
△ The picture shows Rowdy’s company website after using ZMO.AI to generate materials
Nila is the person in charge of a cross-border e-commerce company. Their outdoor sofas are growing rapidly in Europe and the United States, but she also encountered marketing difficulties.
It is a very painful thing to shoot such a large piece of sofa, because not only the transportation cost is high, but also the setting up of the shooting scene is also slow and expensive.
So the Nila team hired many artists to complete the production of materials through P-pictures.
However, headaches also follow. Although the Nila team employs many outsourced retouchers, it often takes more than 10 years of experience to produce very realistic effects, and the level of retouchers varies. The effect and data performance of the P picture are very different from the original picture.
After being introduced by friends in the circle, Nila started to use ZMO.AI’s text P-map. She found that just inputting a paragraph of text, the photo can be modified according to the instructions without any trace of P-picture. It does not require any high-threshold tools or experience. She This Xiaobai can also become a photo master.
The picture produced by P is very natural, and it is completely impossible to see that it has been P, and the data performance is much better than before.
Before using ZMO.AI's products, Nila spent a lot of time every day giving feedback to the retouchers, and it took several days to get better results. "In case the content does not perform well, it needs to be uploaded again, and the time and money consumption here is not small."
△https://background.zmo.ai/
Unlike Nila, Nick is the marketing manager of a professional marketing agency in the United States, responsible for helping advertisers set up official social media accounts and design creative materials.
Nick's customers include both online e-commerce customers, as well as users in traditional industries and even catering industries.
Especially after the epidemic, all businesses cannot do without online marketing, but high-quality materials are indeed a problem.
Nick said so.
The emergence of AIGC has indeed brought great changes to this industry. However, Nick found that the Midjourney or many other AIGC products that are widely rumored on the Internet cannot meet his needs at all, because the details of the products in the generated pictures will change and cannot be 100% maintained. as is.
Nick says:
At first glance, it looks similar, but after careful comparison, it is found that the pattern, logo, and material are not exactly the same, and the merchant will definitely not use the wrong product.
Nick discovered on twitter that the ZMO.AI software can not only fully maintain all the details of the product, but also realistically generate light and shadow, which can fully meet the needs of operators in terms of resolution and realism, which is unique to other AIGC software unattainable.
△https://background.zmo.ai/
Not only that, but what surprised Nick the most was that ZMO's Marketing Copilot function completely redefined the workflow of marketers.
Just upload the product picture, from shooting, to poster, to content optimization, all can be done automatically! Really a mature AI can do marketing by itself, haha
Nick uploaded the client's previous data with good performance to Marketing Copilot to train his own proprietary generative model, so that the output of the model can be more in line with his client's audience preferences and brand tonality.
Marketing Copilot’s model often needs 1-2 weeks of adaptive adjustment of content direction first, and repeats the process of material generation-data feedback-material optimization, and then the unique model automatically generates marketing that suits the audience’s preferences based on instructions and brand tonality Pictures, the feedback from high-quality marketing data is indispensable in this process, and these data are private, and merchants have complete control.
In Nick's view, Marketing Copilot is no longer a simple content generation tool, but a complete set of solutions to change the marketing process. Through the stronger analysis and generation capabilities of AI, it greatly shortens the consumption and collaborative production of all aspects of marketing , and optimize the entire marketing content 24 hours a day based on the final data.
Nick said that the team is indeed considering reducing some marketing personnel after the emergence of AI, because after the team is familiar with the new workflow of Marketing Copilot, the number of images produced for each SKU has skyrocketed from less than 10 to 200 in an instant, and began to use Marketing Copliot conducted a large number of AB tests and iterations, shortening the optimization cycle from three to four months to 2-3 weeks, and sales increased by 3 times.
△https://background.zmo.ai/
It is too early to draw conclusions about what the ultimate AI workflow should look like, but it is likely that it is not just a blank text box, followed by an API that does not belong to you. At present, it is more important for start-up companies to make products that can solve the pain points of merchants and continuously iterate with users, rather than empty talk and iterative demo videos.
Specially optimized large models are needed in specific scenarios
The astonishing growth of ZMO.AI can't help but remind people of TypeFace and Adobe.
Unlike the popular OpenAI, Stability AI and other companies that make general-purpose large-scale models, TypeFace, ZMO, and Adobe all chose the direction of vertical large-scale models with application scenarios for product creation.
There is a common view in the market that all industries in the future will be dominated by very few general-purpose AI models. So is there any further significance for the large vertical model of this application scenario? Obviously, these companies give different views.
Although Typeface is a start-up company established in June 2022, its founder is the well-known former CTO of Adobe—Abhay Parasnis, who also received 6,500 investment from Google Ventures, Microsoft Ventures and Lightspeed at the beginning of its establishment. million dollar investment.
In addition to the star founding team, Typeface is most impressive for large brands, creating personalized text and image content based on corporate brand positioning and audience goals.
Unlike general content generation for the general public, the requirements of big brands for brand tonality and content controllability cannot be achieved by directly calling Stable Diffusion.
Parasnis said:
One of the most fundamental concerns for a company is the security of its data and brand image. Every business wants to make sure they don't inadvertently create inaccurate, plagiarized or offensive content that can tarnish their reputation.
At present, Typeface helps these big brands solve the above pain points through personalized training of brand-specific models and content review algorithms.
Adobe, the originator of the design industry, also released its own AIGC product Firefly some time ago.
Facing the worries of designers losing their jobs, Adobe put forward the slogan of "not replacing, but empowering", and provided practical AIGC functions at a more professional design level.
For example, the function of generating vectors can generate custom vectors with only one sketch, which is very practical for designers.
However, ordinary AIGC generators are often generated as a whole plan, and cannot actually generate vectors and layers.
However, Adobe mentioned at the press conference in October 2022 that it will embed AIGC capabilities into PS, but it has not yet been implemented. Many functions on Firefly are still under development. It is conceivable that Firefly will be integrated into Adobe. It is still a very huge project in the complex tool ecology.
Similar to Typeface and Adobe, ZMO.AI is also a large vertical model for professional users, but ZMO's user group is more of a small B in the marketing direction.
According to Ma Liqian , the co-founder of ZMO.AI , although the basic large models can perform at the level of average humans in many tasks, they do not perform well in specific vertical fields.
This is because domain knowledge in these domains is not common knowledge, and related data are not easily publicly available.
For example, ZMO will need to use a self-developed high-precision matting algorithm in order to fully retain product details. Matting is a complex vision task that involves accurately estimating the alpha value of each pixel to extract foreground objects from images and videos.
This can be challenging due to factors such as complex backgrounds, lighting conditions, and object transparency.
In addition, the labeling of this task is difficult, specific and expensive, and it took ZMO a year and high cost to obtain these high-precision labeling data.
Maliqian said:
In our use case, we might focus on matting a specific object (such as a product), which is not an advantage of a large base model.
It is not difficult to see from the products of Typeface, Adobe, and ZMO that after all, large models need to subdivide the data and rules of the scene. Control input and output can be truly used.
Founder: Image generation quality has reached an inflection point
In order to gain a deeper understanding of the secret behind the crazy growth of ZMO.AI, Zhang Shiying, the founder of ZMO, conducted an exclusive interview. The following is the content of the conversation:
Q1: Over the past year, AIGC has not only attracted many start-up companies, but also many large companies have failed. What do you think are the opportunities for start-up companies?
Zsy: In my opinion, AI companies currently have ecological layers, large model layer, pure application layer and vertical large model layer.
The large model layer is like an operating system. It is an opportunity for a small number of people. It is more suitable for players with strong financial resources and manpower, such as big manufacturers or bosses. Startup companies are more suitable for the latter two types.
The pure application layer has very low barriers in the era when the underlying large model is constantly changing, and the sustainability is not strong. Vertical large-scale models are our firm direction, such companies as Character.AI, Midjourney, Typeface, and ZMO.
Vertical large-scale model companies build an end-to-end engineering stack to cover the entire value chain of model development, training, data, and applications. The products of such companies do not rely on third-party APIs, and they iterate very quickly and can make good use of applications. Data feedback from side users forms a data flywheel.
For example, ZMO fully connects a large number of high-value user feedback data and models on our application side, and uses the data flywheel to guide and optimize the content generation direction of vertical large models, and accumulates its own proprietary data sets.
Q2: Won't vertical large models be replaced by general large models soon? How big is ZMO's model?
Zsy: I don't think that the vertical large-scale model and the general large-scale model are opposite concepts. On the contrary, I think that the vertical large-scale model can stand on the shoulders of the general large-scale model to further optimize the professional field.
For very detailed and personalized C-side scenarios, a general-purpose large model will be more suitable; however, for a very professional toB scenario like marketing, AIGC products will not simply consist of a dialog box and a third-party API behind it.
Because professional scenarios have extremely high requirements for controllability, accuracy, and quality, this will inevitably be a complex system structure, which requires special model optimization on the basis of general large models to meet the needs of marketing scenarios.
Our model parameters are three times that of SD - 2.3B . Of course, we are still using user feedback to optimize RLHF. 50 machines are optimized for training at the same time, maintaining the speed of one iteration per month .
Q4: Midjourney and you have already made a profit. Is Vincento more likely to make a profit than other types of large-scale model companies?
Zsy: I don't know much about other large model fields, so I can't comment. However, for the AIGC field of CV, I think that the generation quality has indeed reached an inflection point, which also explains why the number of paying users continues to grow in large numbers.
For our users, the generated content can reduce their costs and increase their efficiency, and even increase their income. They used to spend tens of thousands of dollars to shoot movies, but now the subscription fee of tens of dollars can be solved, so we The number of paying users quickly rose to 20,000.
Q5: What do you think are the advantages of ZMO?
Zsy: First of all, ZMO has accumulated two and a half years in the field of generative marketing, and has a deep understanding of the know how and user pain points in this field.
For example, users can’t meet the needs of product details, such as the time-consuming and labor-intensive pain points of shooting and PS, etc., so we can build products that deeply bind and market AIGC’s native workflow.
Secondly, we have a large amount of professional data in the marketing field, such as the 60 million high-definition real photo datasets we have accumulated for training ultra-high resolution realistic photos, such as the massive alpha matting datasets we have accumulated, etc.
The last point is that we have built and verified the RLHF feedback system for professional marketing users. This kind of high-quality professional user data feedback is a relatively high commercial data barrier, which belongs to specific industry private data. Only through this feedback can fine-grained parameters Continuously optimize the content generation direction.
Q6: How to deal with the competition from big manufacturers?
Zsy: I think competition is inevitable, but in my opinion, the functions introduced by major manufacturers are more of a defensive response. The applications they make are still carried on the existing business, and they are only aimed at those who have the technology and are willing to pay a high price. A small number of people have made some additional functions; and AIGC startups like us have created products under the framework of the new content creation paradigm from the beginning, building AI Native Apps that are completely different from traditional workflows.
Whether big manufacturers can make powerful new-generation AI products depends on whether they can kill themselves.
Q7: How to achieve user insight?
Zsy: I like chatting with users very much. I spend 2 hours a day reading user feedback. Many users’ pain points are discovered during this process.
I think it is difficult to see the real pain points when looking at the data in the early stage, because there are both our target users and a large number of pure "playing" users in the huge traffic, so finding a large number of target users to observe their use and chatting with them became my goal. It is necessary for daily life. Of course, I also often visit customers' companies and make friends with marketing personnel in different links.