Author丨Linfeng
Editor丨Sea waist
题图丨Midjourney
Last year, Masterpiece Studio, in partnership with NVIDIA, was recommended as one of the best generative AI for 3D modeling, as well as an AI text-to-3D generator that creates 3D avatars, characters, and props that are usable in real time.
Kaedim is also a project launched by NVIDIA's AI Startup Program, located in London, it is an AI 2D-to-3D model company that provides a simple 3D asset creation service for Tusheng, and users can upload 1-6 images to generate 3D models. According to the official website, the platform has generated 51,098 assets, and recently they launched a 3D asset trading platform Marketplace, providing 10,000 easy-to-edit, lossless 3D assets for consumers to choose from. Its main service industry is game entertainment, and it also has partners in e-commerce, architectural design, animation, film and television.
Source: Kaedim's official website
Konstantina Psoma founded Kaedim in 2020 because of her difficult 3D animation for her coursework, and then devoted herself to 2D-to-3D ML algorithms as a graduate student at the University of Bristol in the UK, and founded the company in a campus business competition.
On March 11, Kaedim raised $15 million in Series A led by A16z Games, with participation from Scott Gelb (former president of Riot Games), Nate Mitchell (co-founder of Oculus), Eden Chen (CEO of Pragma), and Siqi Chen (CEO of financial analysis software Runway).
Kaedim has reportedly had 250 large customers, a monthly growth of 20,000 subscribers, and a seven-figure annual revenue. Its C-end users include 3D modeling enthusiasts, interior designers, product designers, and animators.
1. 2D to 3D, to achieve 7-digit annual income
AI generation of 3D is not simple, limited by the difficulty of obtaining and developing training data, although there are companies doing this business, most of them are costly, and it is necessary to think about the commercialization path.
In January of this year, Lumalabs, another 3D asset generation company led by A16z, generated high-quality assets, and the finished model can be "fake and real" in the real world, known as the Midjourney of 3D. In January, Luma AI released Genie 1.0, a text prompt that generates 4 high-fidelity 3D models in 10 seconds.
视频源：Luma官网展示作品，Created by @Jhues
Luma's financing is primarily focused on R&D and innovation. Some analysts pointed out that it is oriented to the C-end, but there is no clear business model, and it is very likely to be commercialized in games, film and television, e-commerce and other scenarios in the future.
Meshy AI is also a highly recommended 3D asset platform, which, like Kaedim, can achieve 3D text and graphics, with fast speed and continuous quality improvement. In the interview, Hu Yuanming, CEO of Meshy and "Chosen Programmer", mentioned that the game has built a bridge for him to start a 3D generative AI, and 3D AIGC is an exciting entrepreneurial direction, although there are many uncertainties, but it is worth exploring with the market. It is not yet known whether the company is profitable.
图源:Meshi i官网
Kaedim's vision is to accelerate the transition from a digital 2D world to a digital 3D world. "Over the past five years, we've seen tremendous growth in 3D technology in markets such as e-commerce, gaming, 3D printing, AR/VR, etc., and we're empowering the next phase of 3D content creation, just as DALL-E helps with 2D graphics," said Konstantina Psoma. ”
They have generated revenue. According to Finsmes, Kaedim has grown to more than seven figures in annual revenue, has more than 250 large clients, and adds 20,000 creators every month. This includes hobbyists and development teams and Fortune 100 companies in industries such as video games, business, architecture, product design, and animation.
Technically, Kaedim uses knowledge such as GAN (Generative Adversarial Network) and transfer learning to build AI products with the help of technologies such as PyTorch and Tensorflow, enabling it to generate and stylize 3D assets. "Humans have seen a lot of cars in their daily lives, and people's vision is in 3D, and when we see an unreleased picture of a new car, we associate it with reality. In the same way, we teach algorithms to recognize 3D data from 2D images. ”
Kaedim says they've helped creators build a complete suite of tools designed to enhance artists' creations, not replace them themselves. They adopt a scheme that integrates tools with workflows such as Unreal and Blender, a set of smart add-ons for 3D modeling pipelines such as automatic UV unwrapping, and a set of managed services that use Kaedim's machine learning models to customize 3D asset generation.
In addition, the process of generating 3D assets from Kaedim diagrams is relatively simple and easy to use, and there are no complicated parameter adjustments.
Kaedim interface
According to the demo video, users can upload 1-6 images of the target generation character in the editing page, select "standard", "high quality", "extreme" three options, corresponding to 1, 2, 4 points of cost, do a good job of naming and height settings, click generate, you can queue up to enter the processing link.
Source: Kaedim Editor Example Model
Source: Model work of DV, a user of the Kaedim Discord community
In addition, users can generate 2D images with prompts, and after getting a satisfactory schematic diagram, click to generate them according to the steps of the previous 3D diagram.
Similarly, in its internal editor, the user can check the skeleton and dimensions of the model, and can also ask the system to enhance the details of the model according to their needs.
Next, the user can use its UV Unwrapping to create a UV map of the model. Professionals can open model wireframes to check the topology and also color models. Models generated from text-to-images can be described or sketched changed in the editor to request additional detail.
Finally, Kaedim allows users to download in the form of image, glb, gitf, obj, etc.
Source: YouTube, Kaedim video presentation
Horizontal comparison of the 3D model effect of the three types of Tusheng, it can be seen that the Kaedim effect is not bad. From the perspective of the white mold effect, Kaedim (Fig. 1) has a clearer line effect, an unknown time, and a strong usability, and supports custom filling. Although Meshy AI (Figure 2) comes with a built-in texture that takes 1 minute, the effect of free generation needs to be enhanced. In addition, Luma's high-quality finished products can also be used directly as game assets, and after using Wall-E prompts to convert to a 3D model, a Wall-E with slightly distorted corners can be generated in about 10 minutes.
Source: Figure 1 is from Kaedim's video demonstration of the Wall-e model, Figure 2 is generated by the author using the image model in Meshy AI, and the original wall-E image is from Wikipedia
Source: Kaedim video demo Wall-e model, Meshy AI integral generation model, Luma free generation model
Recently, Kaedim launched a marketplace for 3D assets with 10,000 assets, which is suitable for indie game developers without bad topology or broken meshes. This site is only accessible to subscribers.
Source: Kaedim
Second, the failure of the department brings entrepreneurial inspiration
On the official website, Kaedim's main visual is a group of cute critters, similar to the temperament of their founder, Konstantina Psoma, a smiling, cheerful, and adventurous Greek girl. She is a young Gen Z entrepreneur who was selected as one of Forbes Europe's "30under30" tech elites last year.
图源:Konstantina Psoma领英
A graduate of the National Technical University of Athens and the University of Bristol, Konstantina left Athens alone at the age of 17 to study in England. Doing 3D AIGC stemmed from her frustration in studying 3D modeling and 3D animation in college. "How hard can this be?" said Konstantina confidently opening Autodesk Maya, a 3D animation software, and finding it frustrated, "It's very hard, and the learning curve is very steep." ”
While other students opted for simple cafes and apartment complexes, Konstantina chose the largest building in town, Bristol Cathedral, which is intricately detailed. The operation was as fierce as a tiger, bothering to render for 3 months, the computer crashed several times, and finally the score failed.
图源：维基百科，图为Bristol Cathedral
"There are hundreds of buttons in Maya, and I'm lost in them. It took me a week to build a chair," Konstantina said, reviewing the process and finding that the process of turning the cathedral into a 3D animation was repetitive and time-consuming, "but there are at least hundreds of 3D objects in a 3D video game in a still frame. ”
Not only that, 3D modeling and design requires mastering most of the software, and the creator can only be handy after having a certain art foundation to operate models, materials, textures, lighting, and rendering. To find an easier modeling "magic", Konstantina spoke to professional practitioners. She found that both 3D designers and game developers revealed a problem: the 3D creation process is a huge bottleneck for the game industry. Even with advanced game engines, developers still need to invest a lot of time and effort in optimizing and tweaking to ensure the quality of the final product.
She can't imagine that in 2019, when most people are talking about VR and AR, countless 3D assets in the game are still modeled by hand. If you can't even generate 3D digital content at scale, how can you achieve a truly digital 3D experience?
A simple entrepreneurial idea came to her mind, "We can help game developers, 3D printing manufacturers, and product designers accelerate their 3D asset production process." Konstantina focused her research on 2D-to-3D deep learning during her graduate studies, where she and her team worked to build machine learning algorithms that convert 2D inputs from sketches and photographs into 3D models. The idea also caught the attention of her mentor, Shawn Layden, the former head of PlayStation's gaming division, and gave him guidance.
In February 2020, Kaedim was founded as a project of NVIDIA's AI Startup Incubation Program.
"A startup is a big idea, and the whole world thinks you're crazy, but you have to be confident. Even though it's just a project, they talk to a number of game studios on their behalf to gain insight into their production processes and see if the product idea fits into the studio's existing work. Konstantina found that modeling is the most labor-intensive part of the process of creating concept art, 3D modeling, and texturing for 3D content, and that 25 to 60 percent of the total cost of game production is spent on 3D modeling.
"Those studios thought, 'You give me a ready-made 3D model, and we'll consider adapting it,'" Konstantina quickly realized that Wensheng 3D and 3D 3D would not work for most game companies, and she chose the 2D to 3D product idea to speed up the studio's production and help them save costs.
The 21-year-old Konstantina proposed a solution that attracted interest from Aardman, the animation company behind Shaun the Sheep. They eventually won the school's new venture competition and successfully secured their first investment, Chris Kingsley, co-founder of British game developer Rebellion Games.
Epic Games, a well-known gaming team, helped develop Kaedim's AI technology in the early stages, and at the same time, Konstantina raised its first funding to continue its venture. In 2021, they launched the first version of the Kaedim web app and started sales, piloting it with 10 companies including Quell (YC 21).
Source: Kaedim
It can be seen that Kaedim was more successful in attracting customers in the early days, but there were twists and turns when launching the product.
After Kaedim opened its beta, thousands of new users joined their Discord, and soon most users questioned that Kaedim was not fully AI-generated and that the quality of AI-generated was low. 404 Media's report that Kaedim's 3D conversion was actually done from scratch by human designers, without the use of any AI — they hired "quality controllers" who were paid $1-$4 per model, and negative reports followed.
Konstantina responded immediately, and quality control was critical to them, creating a feedback loop for the algorithm. Because the output of the AI algorithm is unstable, it needs to be detected by a quality controller to meet the standard before it can be fed back to the user for further editing and adding textures. Her goal has always been to speed up the 3D production process, "[in 2022] the entire process took an average of 15 minutes, which is 10-20 times faster than manual work. As she planned, the process would be shortened to one minute as the algorithm improved.
Addressing the skepticism, Kaedim raised a seed round in October 2022 and secured collaborations with Little Buffalo Studios, 3D printing company Nakkara, gaming or film companies Netflix, Voodoo, Upland, and more, and managed to secure the favor of A16z in 2024, raising the latest round of $15 million. A16z calls it "the next generation of co-development platforms for the 3D world." ”
According to Konstantina, as the concept of the metaverse and digital space becomes more popular, the demand for 3D modeling will only increase. At present, the company's employees have increased from 2 to 50 at the beginning, and the number of employees is still increasing.
3. Speed up game developers by 10 times
The company doesn't offer a free trial, and you pay for all the steps involved in coloring and reshaping the model, starting at $300/month. This is also a point where Kaedim has been complained about a lot, but it does not delay most 3D designers from exploring it.
Source: Kaedim
In its Discord community, there are a large number of users showing the effect of the model in the "show your model" thread. For example, active user liamine created a "Gundam" series, and he used Kaedim 2D rendering to generate 3D modeling, and Blender 3D animation, texture, and rendering with high-quality results.
Source: Discord user Liamine Thread Page Share
Looking at the specific effect of the collaboration, Kaedim provides a solution to a dilemma known to the designer or developer.
"We're 10 times faster for game developers to create pipelines for 3D assets. The Kaedim team said. Gaming is the main application industry for Kaedim's generated assets, and they have served AZRA, Myth, Upland, Voodoo (a well-known French casual game company), Voldex, etc.
Source: Kaedim
The first is to bridge the gap in time, resources, and production difficulty. The industry consensus is that the 3D workflow is long, the modeling cost is high, and the asset production is slow. It has been revealed that even a small production of 3-minute 3D animation costs hundreds of thousands of yuan or more. However, due to factors such as scale, project architecture, and production cycle, the cost of 3D production cannot be specific to the accurate market price.
Kaedim sees this and will help animation studios fill the gap in 3D assets. Myth is a London-based animation studio that has cooperated with Fanta, Pixelart, etc., and mainly takes orders for 2D animation businesses such as basic graphics and illustration animation. According to reports, their designers spend a lot of time creating 3D animation assets in Cinema 4D, often taking a day or half a day, but with Kaedim, they can control the time to 15-30 minutes, and create multiple grids in a few hours, "We can make a new motion test/short film every week. ”
Source: Kaedim, a short film with whom Myth collaborated
The second is to provide an asset creation pipeline that streamlines the development process and creates a seamless bridge. The all-time Cyberpunk 2077 cost $450 million, and updates and patches alone cost $41 million. According to GameLook, in the field of game development, the art department will play the role of "gold eater", and in 3D open world and 2D card games, art often accounts for more than 70% of the core development team. The game field has high requirements for 3D assets, and even with the advent of the AIGC era, batch generation of assets with rough modeling, lack of detail, and distorted structure cannot improve efficiency.
That's why Kaedim is integrating AI into workflows to increase efficiency, as in its partnership with Upland. Upland is a blockchain virtual property trading game with more than 4 million virtual assets, and Upland is also cultivating a UGC community, in order to promote the game experience, they have requirements for the number and speed of 3D asset creation. According to reports, Kaedim has added features such as automatic texturing, UV unwrapping, and LoD (Level of Detail) for them.
Source: Kaedim, an example of a 3D asset in collaboration with Upland
"Kaedim has helped us scale up the production of in-game 3D assets while expanding the UGC platform. According to Upland's case, Kaedim provided a solution for its image-to-mesh technology, eliminating the steps of manual grayscale modeling. The 70 models that would have taken 245 days (non-AI) to produce were completed in just 70 days, using an AI-driven 3D process to reduce the production time of 5,000 featured totem life forms by 72%.
But in terms of how it works, Kaedim emphasizes that the company combines AI and human expertise, and does not rely entirely on AI generation. Originally, in August 2022, Kaedim's external promotion was "generating customized 3D models in minutes" and "automated 3D modeling", etc., as mentioned above, after the media revealed in September 2023 that Kaedim would have human artists for quality control, the official website responded and corrected the statement.
In a post on Medium, Konstantina Psoma said that human quality control helps train algorithms: "It's very important for the company that the 3D team members work with machine learning engineers to provide technical feedback. In addition, they work with 3D designers to optimize available, ready-to-invest 3D assets to ensure that the output of 3D assets meets industry standards.
Over the past year, Kaedim has released a suite of automation tools including auto-texturing, auto-UVs, and generating LoD (Level of Detail) capabilities, adding 3D asset "pre-modifiers" for customers to customize workflows. In addition, Kaedim's target industries are not only games, but also the asset needs of companies such as interior design and e-commerce.
"We still have a long way to go in the development of our technology to realize the 3D modeling workflow of the future that we envisioned. We will always focus on the customer, work hard, and keep building. ”
Although Konstantina later learned that there really wasn't a quick recipe for 3D modeling, she was shaping the platform to speed up the modeling process so that people only had to say a magic mantra when they thought of generating 3D assets quickly: Kaedim.