On May 23, 36Kr held a "Subversion· AIGC" Industry Development Summit. This summit brings together industry forces, discusses the coping strategies of enterprises and industries in the face of change, shares thinking, explores and discovers the most potential enterprises and the most valuable technologies in the industry, and explores the way forward in a turbulent environment.
During the conference, He Wanyu, CEO of Xiaoku Technology, delivered a keynote speech entitled "AIGC Creation and Sharing Road in the Field of Pan-architectural Design".
He Wanyu, CEO of Xiaoku Technology, delivered a keynote speech entitled "AIGC Creation and Sharing Road in the Field of Pan-architectural Design"
He Wanyu first explained what is the "field of pan-architectural design", she introduced: pan-architectural design is a very broad field, including urban planning, architectural design, garden landscape and interior decoration. The total market size is about 910 billion in China and about 6.7 trillion in the world, which is a very huge market.
Xiaoku Technology believes that artificial intelligence plus people must be greater than the existence of people themselves. Under such a path, for the field of pan-architectural design, AI can help people improve quality and efficiency simultaneously.
He Wanyu prospected: The trillion-level pan-architectural design field (planning, architecture, interior, landscape, etc.) urgently needs to improve quality and efficiency tenfold. This depends on AI+ people > people. Xiaoku AI Cloud One-stop Pan-construction AIGC Innovation and Sharing Platform is a self-developed large model based on tens of millions of pan-building data, which accurately and controllably generates design intentions to achieve "zero threshold generation", "barrier-free alchemy" and "seamless workflow".
The following is the transcript of the speech of He Wanyu, CEO of Xiaoku Technology (excerpted by 36Kr):
Good afternoon everyone!
It is a great honor to have the opportunity to share with you some of our creative paths about AIGC in the field of pan-architectural design, and the previous instant design friends talked about a lot of useful explorations in the field of interaction design.
And what is the pan-construction field? Pan-architectural design is a broad field, including urban planning, architectural design, landscape and interior decoration. The total market size is about 910 billion in China and about 6.7 trillion in the world, which is a very huge market.
How do we work in this trillion-dollar market? In fact, it is completely dependent on people, relying on people's experience, people's time, plus a series of traditional tools to achieve. In this field, there are also many words such as "volume", "consumption", "boiling". Because the work of pan-architectural designers is very heavy, not only need to systematically design an idea, but also use a drawing to implement it, and there is more work on how to gradually transform these things into a combination with the real world.
To solve the core pain points of designers, the key is to solve quality and efficiency, and in the case of enhanced quality and efficiency, we do not increase costs, or even reduce costs. But how to achieve 10x or even 100x improvement in quality and efficiency? This is the core topic of our industry's next phase of upgrading.
How is this upgrade achieved? It's actually through artificial intelligence. We believe that artificial intelligence plus people must be greater than the existence of people themselves. Under such a path, for the design of the pan-architectural field, AI can help people improve quality and efficiency simultaneously.
Xiaoku Technology was founded in 2016, and in 2017, we were already exploring how to use convolutional neural networks to generate architectural planning designs on maps. For example, how each plot is generated by AI, and the results can be further edited and output; In 2018 and 2019, we also began to use adversarial networks to generate a series of architectural intentions, and we can also automatically design and generate house types. In 2019 and 2020, we further explored how to combine graphics and images, where a person sketches on the left and a machine on the right can generate architectural intentions, when images were still blurry. From 2020 to 2021, the image will be used to generate graphics, which is accurate and editable, and can also further generate three-dimensional models, in this case, we will generate knowledge and AI in the field of architecture, as well as image model control, and further realize the application in the industrial field.
In 2022, we will begin to generate architectural images with text, combining the database we have accumulated over the years, especially Chinese corpus, architectural label labeling, etc., with image data, which can generate architectural intentions in Chinese context. For example, we talked about a master architect, Zaha Hadid, who represents the logic of future style architecture. But generic large models don't respond to the term, but with our vertical data training, such a futuristic architectural style can be generated. Therefore, it can be seen that in addition to large models, AIGC has become particularly important in each vertical field.
Our small library has 7 years of accumulation of ARCHINET as the most important database, which consists of tens of millions of pan-architectural data, including the urban planning, architectural design, landscape architecture, interior decoration we just mentioned a series of fields. It has structured labeled data and Chinese corpus, on this basis, we train a large amount of computing power and optimize at the level of deep learning operators to achieve three of our own image large models: small library building large model, indoor large model, pan-architectural all-encompassing large model.
We compared the self-developed large model with some oversimulations and models that failed experimentation, such as for this set of prompts: "residential, high-rise elevations, modernist style". In fact, this overfitting model has similar effect to our generation under this set of inputs, because the overfitting model can directly generate something very accurate. But look at the back, if you add new words, the overfitting model does not react, but generates something with similar effect each time; And look at the small library model, if you enter new Chinese style or people diagram, semi-bird's eye view, future style and other words, you will have a series of reactions to such new input inscriptions. This is the difference between vertical domain models and overfitting models in terms of specialized details.
Therefore, based on the small library big model, we created a new product, which was released in early May this year, called Xiaoku AI Cloud, a one-stop pan-construction AIGC innovation platform. It includes three major sections: generation, training, and sharing.
Small library AI cloud
- The first core content "generation" has two modes, one is to generate pictures through text, and the other is to generate further images through pictures and graphics, which are more common and have unique advantages in our field;
- The second piece of "training", we know that in the field of architecture or in the field of vision, in addition to direct generation, we can also install a system, deploy the environment with their own data set training, but this threshold is very high, for ordinary designers it is basically impossible to build such a complex environment. For most people, not only hope to use generation, but also hope to be able to train their own models, further use their own work field, we also give model training capabilities here;
- The third block "Sharing", users can share training results and generated results, and publish and share them on our platform.
For example, in a typical workflow, we can provide a lot of feedback on different building types and requirements, and the entire system is used in the cloud, as long as it is accessed through clould.xkool.ai. In the "Generative" module, AI-generated results would have required the architect's unique design and aesthetic capabilities, which would have taken several days. Here, you only need to call the large model, enter the architectural vocabulary related to the demand, and get unexpected results in 10 seconds. An important detail here can be noted, whether the lines of the building are straight or not, is a very important sign of the usability of the model, not very confusing and blurry things, is a very accurate output.
There is also an important feature here that distinguishes it from existing overseas products: LoRA overlay. Different styles can be realized through various LoRA small models, and users can train themselves or select other people's models for calling. For example, handmade model style, or acrylic model style. Through the "Tusheng Chart", you can upload a reference map, copy the prompters and parameters of the picture we see in the inspiration square, and directly copy the data into the input, automatically generating a series of outputs very similar to the reference just now. For example, upload a photographed indoor rough space, enter the requirements, and you can also generate a series of results for this photo.
For designers, more scenes are precisely generated and rendered, for example, I want to turn a hand-drawn sketch into a specific architectural style. You can upload a sketch, enter requirements, and then generate them. You can get very precise and clear results in 3 simple steps, based on sketches. At the same time, we can also edit the generated images, for example, I am not satisfied with the entrance of this building and want to make a series of changes to the entrance. We can directly paint the space of this part of the entrance, and then make a new generation of the entrance, enter your needs, and further generate the entrance effect you want. This process can be carried out continuously, and here we can also change the entrance to wood again and see what kind of effect it will produce? If you are not satisfied with the entire building, but just think that the surrounding environment is not bad, you can erase the main building and replace it with a scheme that may be more different, such as a more curvy futuristic building, which can also be tested immediately.
The above series of actions are based on the creative scenes of pan-architectural designers. In addition to generation, you can also use image materials to "train" the model to do Fine Tuning. You only need to upload the image data of the building class, do a simple parameter setting and start training. Maybe it took you 10 minutes to collect materials, 1 minute to upload and set up, and finally you only need to go through another 10 minutes of ultra-fast cloud training, and you can immediately get a model of your own.
After generating results and creating tools, designers can do further sharing and sharing. We are constantly having user-generated and shared content in the Inspiration Square and Model Marketplaces. Ordinary users can also see what your build parameters look like, and they can learn to improve their control over the build quality. Or directly call the model uploaded or generated by other alchemists to make such a model.
In the Inspiration Square, you can not only watch what others do, but also generate the same model with one click, and you can get content that is similar to others but not exactly the same. In the model bazaar, users can call the models in it, and we can choose different LoRA weight ratios to achieve their own results through the superposition of different small models.
We also did a lot of testing. For example, the same set of input prompts compared to MidJourney. It can be seen that the following MJ effect is actually very exaggerated for architecture, and this color and form are far from the effect that pan-architectural designers want in real work. Or another set of tivotions, the effect of the curve is much worse. It can be seen that the MJ colors below are relatively close to the dream two-dimensional effect, and the effect of the small library above is the style and relatively unified volume relationship preferred by pan-architectural designers. [Animated Emoticon]
Comparison between Xiaoku AI Cloud and MidJourney
For example, we also compared it to Stable Diffusion. For example, the same set of inputs as just now, because the underlying layer of SD is very basic, basically can not achieve any desired generation results. But in the field of pan-architecture, our own large model has done a lot of pre-training for architecture, and we can see that different teleprompting feedback is not very accurate at the entire level of SD, but the effect generated by the small library is very accurate to express what the user wants here.
In summary, the generation and training level of Xiaoku absorb the advantages of foreign AIGC products, and effectively avoid their disadvantages, and become a product that allows pan-architectural designers in China to produce high-quality results with low threshold and high efficiency.
Comparison of Xiaoku AI cloud products
What else will we do in the future?
Because Xiaoku has accumulated AI capabilities for various design links in the construction field in the past, we are organically combining Xiaoku's tool matrix with Xiaoku's AI cloud, which does not refer to blunt jump links, but will use the new ChatBot to mobilize Xiaoku's AI capability APP to rebuild and complete a series of workflows.
Such as talking to a small library robot, such as entering the demand for a house in Guangzhou. AI will give the proposed apartment type, and then call the "smart single" APP in the "Xiaoku Design Cloud", which can edit and design the apartment type and further generate a three-dimensional model. Through this model, we can call the generation function of the small library AI cloud to further generate the solution we want on this model. Furthermore, corresponding CAD drawings and 3D models can be generated.
This series of work took less than an hour, and the entire workflow, which used to take about 10 days and multiple different teams to implement, was reshaped by Xiaoku AI. Not only can designers repeatedly deliberate on it to find the best solution, but it can also make design and communication very efficient with the help of machines. At the same time, the result can be transformed from an intention into a real design plan.
We have more application space in other pan-construction fields, and hope to have the opportunity to explore together with more friends. For example, in the indoor field, there is an indoor blank effect, and the interior decoration effect is generated immediately after taking pictures. Or after uploading the floor plan, on the one hand, the apartment type can be evaluated, and on the other hand, a series of interior design effects can be generated that the user wants.
At present, Xiaoku AI cloud is the concept of design workflow, because AIGC's output in the pan-construction field is more conceptual, morphological, and modeling style content. But when this ability is combined with other capabilities of Xiaoku: graphic design, model design, architectural expertise and other fields, we can already see that Xiaoku Design Cloud is a platform that can cover the whole process of the entire field of pan-architectural design. The results it produces can be connected to the pre-planning stage, post-evaluation, or even implementation stage. So that the results of Xiaoku AI Cloud are no longer just a scheme or idea in the pan-construction field, but a design scheme that can truly go from a scheme idea to a practical application.
The above is our creative exploration of AIGC in the field of pan-architectural design. In summary, Xiaoku AI Cloud is a one-stop AIGC creation and sharing platform with zero threshold generation, barrier-free alchemy, and no connected workflow.