News of july 3rd, this morning, mEta released their vincent 3d model mEta 3d gen (known as 3dgen) dropped a heavy bomb on vincent's 3d track。
3dgen was able to quickly generate 3d assets (meaning digital resources for the production of 3d animations, such as 3d models, texture maps, materials, etc.) over a period of one minute, 3-10 times faster than other systems available. It can also quickly adjust texture maps based on the same 3d shape, helping creators to achieve rapid iterativeity。
Pbr is the most important feature of 3dgen. Most of the previous 3d generation models did not have this function, which made the practical application of 3d assets generated by ai unconstructive。
Pbr can integrate the physical properties of the 3d model into 3d assets, making the 3d model interact more natural with the simulated physical world, and is an important technique for creating a veritable image and natural reality. 3dgen, which supports pbr, is better integrated into the work stream of game developers and late film producers。
MThe technical methods behind 3dgen are also described in detail in the technical documentation. They have adopted a two-step approach to generating 3d assets. One model, assetgen, was used to generate the base 3d shape and texture, while another model, texturegen, was designed to produce higher quality textures and optimize the results of assetgen。
3dgen presentation video (source: m)I don't know
3dgen on the technology pathThe tests carried out by eta were highly valued by professional creators and ordinary people, and scored higher than most 3d-generated models in terms of the loyalty of the hint, texture quality, geometry and image quality。
I. One minute to generate high-quality 3d assets, 3-10 times faster than other existing systems
Production of 3d assets has been one of the longest-consuming and most challenging issues in the design and development of video games, vr/ar applications and the film special effects industry. MThe present launch of 3dgen is intended to provide a possible solution to this problem。
MIn its technical report, eta stated that 3d asset generation was far more difficult than image and video generation. The 3d assets in practical application are extremely demanding and are not currently available for training. Data for image and video modelling training is in the billions, while the number of 3d assets available for training is three to four orders of magnitude. Therefore, 3d generation must learn from non-3d images and videos and learn to extrapolate 3d information from 2d observations。
3dgen supports complex text tips, which are more faithful than the other authors ' 3d methods. MEta claims that 3dgen generated a 3d asset in less than one minute, 3-10 times faster than other existing systems。
After the initial generation has been completed, the 3dgen can also reorder the texture of the original 3d asset within 20 seconds, depending on the needs given by the new hint。
MWhile eta does not currently publicly test the system, they publish a three-minute video on social media x that shows the results and functions of the model。
The phrase below which the result is generated is “a pug made out of mEtal. The results are fully consistent with the textual content, and the structure of the 3d grid (the framework of the 3d model) of the metal huba dog in the picture is consistent with the appearance of the haba dog in the real world, and the texture clarity and resolution are desirable。
3dgen generates 3d assets of high texture quality (source: m)I don't know
MIn a video shared on x, eta also shows 3dgen's texture adjustment function. This function allows 3dgen to change the texture of the whole scene by the given style. In the presentation video below, the tester entered terrorist films, pixels, and 3d textures of a given style were quickly generated by 3d. These textures are applied uniformly to all 3d images in the scene and produce better quality。
3dgen gives different textures to the same 3d grid according to the hintI don't know
Ii. Support for physical-based rendering and manual 3d grid input that is more easily integrated into the creator's workflow
In the past, ai generated 3d assets that generally did not have real light and material properties, thus limiting their usefulness in professional workflows. The 3dgen system supports the physical-based rendering (pbr) and human artist 3d grid input, which allows creators to integrate 3d assets generated by 3dgen more easily into their own work streams。
The pbr technology integrates properties based on physical patterns into 3d maps, such as the albedo of a particular material, roughness, metallicity, opacity, refractive effects, spontaneous light, etc。
The application of this technology allows computers to understand not only the shape of the 3d asset, but also how it interacts with factors in the physical world such as light, which is essential for creating a veritable image and a natural reality environment, as well as ensuring consistency in the creative process。
This technology is now widely used in the areas of architectural visualization, games, vr/ar, film specialty production and is an important means of creating high-quality 3d assets. The 3d image generated by the 3dgen, which supports the pbr technology, not only has a high quality in static conditions, but also maintains a high degree of authenticity when the 3d image moves。
On the mThey typed "a futuistic robot." this hint and allowed mixamo, a 3d computer graphics technology company, to complete the auto-coupling and animation of the characters of the 3d asset. In this picture, the luminous changes in surface metals when robots dance are real, thanks to pbr technology。
A moving 3dgen generation result (source: m)I don't know
3dgen supports 3d grid input by human artists, which is a practical feature. In the long-term process of creation, artists or businesses have accumulated a large number of artificial 3d grids, the quality of which is generally higher than that of ai. 3dgen can use ai to give the grid a variety of textures, based on the 3d grid produced by human artists, while maintaining relative stability in the quality of the results produced at a fast pace。
The image below is a texture 3d grid produced by a human artist。
No texture 3d grid (source: m)I don't know
Enter "a majestic mo"Narch butterfy and "a mighty butterfly, arcane sigils on its wings, sparkling." after these two hints, 3dgen can make this non-articular butterfly live。
3d assets based on a textless 3d grid (source: m)I don't know
Iii. Quality 3d generation achieved by two key components, loyalty and texture quality well above other models
MTea presented the technical rationale of 3dgen. They divided the 3d generation into two steps and used two components, one for text to 3d and the other for text to text to text, a method that significantly enhanced the 3d generation。
Model m for text to 3d generationEta 3d assetgen (known as assetgen), which is primarily used to create the initial 3d asset, i. E. A 3d grid with rougher textures and pbr materials (like the non- textureed butterfly shown above). The process took only 30 seconds。
Model m for text to text generationEta 3d texturegen (known as texturegen), which produces better textures and pbrs for this 3d asset, or creates textures of different styles. This process can be completed in only 20 seconds。
Technological rationale for 3dgen (source: m)I don't know
Assetgen and texturegen allow 3dgen to merge images, 3d shapes and appearances, and uv textures. Assetgen completed a more complex 3d shape generation for texturegen, a model that optimizes uv texture。
Texturegen's training data are uv textures for human artists, and these high-quality training data significantly enhance model performance and therefore produce more authentic textures than assetgen. Assetgen and texturegen can be said to be complementary models. Besides, mEta says the entire process of generating this 3d asset is based on mOn top of emu。
To assess the effects of 3dgenTea invited professional creators and ordinary people to participate in their research, and graded hints for loyalty and visual quality。
MEta selected several currently accessible 3d generation models for comparison. 3dgen scored more than other models at both the 3d shape generation and the texture generation stages. 3dgen ranked first in the scores by type of object generated, in object and combination scenes, and slightly less than in other models in character。
3dgen hint loyalty score (source: m)I don't know
Quality of generation, mThe eta selected three better performing third-party models, allowing professional creators and ordinary people to evaluate the results of model generation. The results show that 3dgen has an overwhelming advantage over loyalty, image quality, texture quality and geometry。
3dgen produces quality compared to other models (source: m)I don't know
MAfter official eta announced 3dgen, netizens immediately came up with many possible applications. There are a number of netizens who think it would be a good attempt to print out the 3d asset generated, and those who believe that it will radically change the rules of the game development industry。
Netizen evaluation (source: x)
Conclusion: 3d generation technology re-engineered, with conventional 3d modelers bearing the brunt
The rapid development of ai 3d generation technologies is important for achieving technological advancements in the 3d modelling industry. Manual 3d modelling is time-consuming and labour-intensive, and the use of 3d images in games and video works was previously the prerogative of a small number of large companies, and ai 3d generation technology could simplify 3d modelling processes and reduce 3d modelling time。
Currently, a number of industries have used 3d generation techniques. The free 3d model meshy, published by the american 3d generating enterprise meshy ai, is now being used by a large number of independent game developers, allowing small teams of people to develop very real games。
However, the emergence of this technology also raises questions about the future of the 3d modeling profession. While 3dgen may increase the productivity of 3d artists, it may also reduce the demand for the creation of certain types of conventional 3d assets. Perhaps in the future, creators need to pay more attention to creative directions that are difficult for ai to replicate or to engage in more sophisticated customization。
Source: mI'm not sure i'm going to do this
