Home Technology peripherals AI Based on DiT and supporting 4K image generation, Huawei Noah 0.6B Vincent graph model PixArt-Σ is here

Based on DiT and supporting 4K image generation, Huawei Noah 0.6B Vincent graph model PixArt-Σ is here

Mar 11, 2024 pm 01:07 PM
industry

This model uses the DiT framework like Sora.

As we all know, developing top-level T2I models requires a lot of resources, so it is basically impossible for individual researchers with limited resources. Affordability has also become a major obstacle to innovation in the AIGC (artificial intelligence content generation) community. At the same time, as time goes by, the AIGC community will be able to obtain continuously updated, higher-quality data sets and more advanced algorithms.

So the key question is: how can we efficiently integrate these new elements into the existing model and make the model more powerful with limited resources? ?

In order to explore this problem, a research team from research institutions such as Huawei’s Noah’s Ark Laboratory proposed a new training method: weak-to-strong training (weak- to-strong training).

Based on DiT and supporting 4K image generation, Huawei Noah 0.6B Vincent graph model PixArt-Σ is here

Paper title: PixArt-Σ: Weak-to-Strong Training of Diffusion Transformer for 4K Text-to-Image Generation
Paper address: https: //arxiv.org/pdf/2403.04692.pdf
Project page: https://pixart-alpha.github.io/PixArt-sigma-project/

Their research is based on PixArt-α, an efficient Vincentian graph training method they proposed last October. Please refer to the report on this site "The ultra-low training cost Vincentian graph model PixArt is here, and the effect is comparable to MJ. It only needs SD 10% training time》. PixArt-α is an early attempt at the DiT (Diffusion Transformer) framework. Now, with Sora on the hot search and Stable Diffusion emerging in endless applications, the effectiveness of the DiT architecture has been verified by more and more work in the research community, such as PixArt, Dit-3D, GenTron, etc. [1].

The team used the pre-trained base model of PixArt-α and integrated advanced elements to promote its continuous improvement, ultimately resulting in a more powerful model PixArt-Σ. Figure 1 shows some examples of generated results.

Based on DiT and supporting 4K image generation, Huawei Noah 0.6B Vincent graph model PixArt-Σ is here

How to make PixArt-Σ?

# Specifically, in order to achieve weak-to-strong training and create PixArt-Σ, the team adopted the following improvement measures.

Higher quality training data

The team collected a high-quality data Set Internal-Σ, which mainly focuses on two aspects:

#(1) High-quality images: This dataset contains 33 million high-resolution images from the Internet, all Over 1K resolution, including 2.3 million images at approximately 4K resolution. The main features of these images are their high aesthetics and cover a wide range of artistic styles.

(2) Dense and accurate description: In order to provide a more precise and detailed description for the above image, the team replaced the LLaVA used in PixArt-α with a A more powerful image descriptor Share-Captioner.

Not only that, in order to improve the model’s ability to align text concepts and visual concepts, the team extended the token length of the text encoder (i.e. Flan-T5) to approximately 300 words . They observed that these improvements effectively eliminate the model's tendency to hallucinate, enabling higher quality text-image alignment.

Table 1 below shows the statistics of different data sets.

Based on DiT and supporting 4K image generation, Huawei Noah 0.6B Vincent graph model PixArt-Σ is here

Efficient token compression

To enhance PixArt-α, the team The generation resolution has been increased from 1K to 4K. In order to generate ultra-high resolution (such as 2K/4K) images, the number of tokens will increase significantly, which will lead to a significant increase in computing requirements.

To solve this problem, they introduced a self-attention module specially tuned for the DiT framework, which uses key and value token compression. Specifically, they used grouped convolutions with stride 2 to perform local aggregation of keys and values, as shown in Figure 7 below.

Based on DiT and supporting 4K image generation, Huawei Noah 0.6B Vincent graph model PixArt-Σ is here

In addition, the team adopted a specially designed weight initialization scheme to achieve smooth adaptation from pre-trained models without using KV (key-value) compression. This design effectively reduces training and inference time for high-resolution image generation by approximately 34%.

Weak to strong training strategy

The team proposed a variety of fine-tuning technology that can quickly and efficiently adjust weak models to strong models. These include:
(1) Replacement uses a more powerful variational autoencoder (VAE): PixArt-α's VAE is replaced by SDXL's VAE.

(2) From low resolution to high resolution expansion, in order to cope with the problem of performance degradation, they use the Position Embedding (PE) interpolation method.

(3) Evolve from a model that does not use KV compression to a model that uses KV compression.

#The experimental results verified the feasibility and effectiveness of the weak-to-strong training method.

Through the above improvements, PixArt-Σ can generate high-quality 4K resolution images with the lowest possible training cost and as few model parameters as possible.

Specifically, by starting with fine-tuning from an already pre-trained model, the team only used an additional 9% of the GPU time required by PixArt-α. A model capable of generating 1K high-resolution images. This performance is remarkable because it also uses new training data and a more powerful VAE.

In addition, the parameter amount of PixArt-Σ is only 0.6B. In comparison, the parameter amount of SDXL and SD Cascade are 2.6B and 5.1B respectively.

The beauty of the images generated by PixArt-Σ is comparable to that of current top-notch PixArt products, such as DALL・E 3 and MJV6. In addition, PixArt-Σ also demonstrates excellent capabilities for fine-grained alignment with text prompts.

Figure 2 shows the result of PixArt-Σ generating a 4K high-resolution image. It can be seen that the generated result follows complex and information-dense text instructions well.

Based on DiT and supporting 4K image generation, Huawei Noah 0.6B Vincent graph model PixArt-Σ is here

Experiment

##Implementation details

Training details: For the text encoder that performs conditional feature extraction, the team used a T5 encoder (i.e., Flan-T5-XXL) following the practices of Imagen and PixArt-α. The basic diffusion model is PixArt-α. Different from the practice of extracting a fixed 77 text tokens in most studies, the length of text tokens is increased from 120 in PixArt-α to 300 because the description information organized in Internal-Σ is more dense and can provide highly fine-grained details. . In addition, VAE uses a pre-trained frozen version of VAE from SDXL. Other implementation details are the same as PixArt-α.

The model is fine-tuned based on PixArt-α’s 256px pre-training checkpoint and uses positional embedding interpolation technology.

The final model (including 1K resolution) was trained on 32 V100 GPUs. They also used an additional 16 A100 GPUs to train 2K and 4K image generation models.

Evaluation indicators: In order to better demonstrate the aesthetics and semantic capabilities, the team collected 30,000 high-quality text-image pairs to compare the most powerful text-image pairs Models are benchmarked. PixArt-Σ is primarily evaluated here by human and AI preferences, as the FID metric may not appropriately reflect the generation quality.

Performance comparison

Image quality assessment: The team qualitatively compared PixArt- Σ versus closed-source T2I products and open-source model generation quality. As shown in Figure 3, compared to the open source model SDXL and the team’s previous PixArt-α, the portraits generated by PixArt-Σ are more realistic and have better semantic analysis capabilities. PixArt-Σ follows user instructions better than SDXL.

Based on DiT and supporting 4K image generation, Huawei Noah 0.6B Vincent graph model PixArt-Σ is here

#PixArt-Σ not only outperforms open source models, but is also competitive with current closed source products, as shown in Figure 4.

Based on DiT and supporting 4K image generation, Huawei Noah 0.6B Vincent graph model PixArt-Σ is here

Generate high-resolution images: The new method can directly generate 4K resolution images without any post-processing. In addition, PixArt-Σ can also accurately comply with complex and detailed long text provided by users. Therefore, users do not need to bother designing prompts to get satisfactory results.

Human/AI (GPT-4V) preference study: The team also studied human and AI preferences for generated results. They collected the generation results of 6 open source models, including PixArt-α, PixArt-Σ, SD1.5, Stable Turbo, Stable XL, Stable Cascade and Playground-V2.0. They developed a website that collects human preference feedback by displaying prompts and corresponding images.

#Human evaluators can rank images based on generation quality and how well they match the prompt. The results are shown in the blue bar graph of Figure 9.

It can be seen that human evaluators prefer PixArt-Σ to the other 6 generators. Compared with previous Vincentian graph diffusion models, such as SDXL (2.6B parameters) and SD Cascade (5.1B parameters), PixArt-Σ can generate higher quality and more consistent with user prompts with much fewer parameters (0.6B) image.

Based on DiT and supporting 4K image generation, Huawei Noah 0.6B Vincent graph model PixArt-Σ is here

Additionally, the team used an advanced multi-modal model, GPT-4 Vision, to perform AI preference studies. What they do is feed GPT-4 Vision two images and let it vote based on image quality and image-text alignment. The results are shown in the orange and green bars in Figure 9, and it can be seen that the situation is basically consistent with human evaluation.

The team also conducted ablation studies to verify the effectiveness of various improvement measures. For more details, please visit the original paper.

Reference article: 1. https://www.shoufachen.com/Awesome-Diffusion-Transformers/

The above is the detailed content of Based on DiT and supporting 4K image generation, Huawei Noah 0.6B Vincent graph model PixArt-Σ is here. For more information, please follow other related articles on the PHP Chinese website!

Statement of this Website
The content of this article is voluntarily contributed by netizens, and the copyright belongs to the original author. This site does not assume corresponding legal responsibility. If you find any content suspected of plagiarism or infringement, please contact admin@php.cn

Hot AI Tools

Undresser.AI Undress

Undresser.AI Undress

AI-powered app for creating realistic nude photos

AI Clothes Remover

AI Clothes Remover

Online AI tool for removing clothes from photos.

Undress AI Tool

Undress AI Tool

Undress images for free

Clothoff.io

Clothoff.io

AI clothes remover

Video Face Swap

Video Face Swap

Swap faces in any video effortlessly with our completely free AI face swap tool!

Hot Tools

Notepad++7.3.1

Notepad++7.3.1

Easy-to-use and free code editor

SublimeText3 Chinese version

SublimeText3 Chinese version

Chinese version, very easy to use

Zend Studio 13.0.1

Zend Studio 13.0.1

Powerful PHP integrated development environment

Dreamweaver CS6

Dreamweaver CS6

Visual web development tools

SublimeText3 Mac version

SublimeText3 Mac version

God-level code editing software (SublimeText3)

DeepMind robot plays table tennis, and its forehand and backhand slip into the air, completely defeating human beginners DeepMind robot plays table tennis, and its forehand and backhand slip into the air, completely defeating human beginners Aug 09, 2024 pm 04:01 PM

But maybe he can’t defeat the old man in the park? The Paris Olympic Games are in full swing, and table tennis has attracted much attention. At the same time, robots have also made new breakthroughs in playing table tennis. Just now, DeepMind proposed the first learning robot agent that can reach the level of human amateur players in competitive table tennis. Paper address: https://arxiv.org/pdf/2408.03906 How good is the DeepMind robot at playing table tennis? Probably on par with human amateur players: both forehand and backhand: the opponent uses a variety of playing styles, and the robot can also withstand: receiving serves with different spins: However, the intensity of the game does not seem to be as intense as the old man in the park. For robots, table tennis

The first mechanical claw! Yuanluobao appeared at the 2024 World Robot Conference and released the first chess robot that can enter the home The first mechanical claw! Yuanluobao appeared at the 2024 World Robot Conference and released the first chess robot that can enter the home Aug 21, 2024 pm 07:33 PM

On August 21, the 2024 World Robot Conference was grandly held in Beijing. SenseTime's home robot brand "Yuanluobot SenseRobot" has unveiled its entire family of products, and recently released the Yuanluobot AI chess-playing robot - Chess Professional Edition (hereinafter referred to as "Yuanluobot SenseRobot"), becoming the world's first A chess robot for the home. As the third chess-playing robot product of Yuanluobo, the new Guoxiang robot has undergone a large number of special technical upgrades and innovations in AI and engineering machinery. For the first time, it has realized the ability to pick up three-dimensional chess pieces through mechanical claws on a home robot, and perform human-machine Functions such as chess playing, everyone playing chess, notation review, etc.

Claude has become lazy too! Netizen: Learn to give yourself a holiday Claude has become lazy too! Netizen: Learn to give yourself a holiday Sep 02, 2024 pm 01:56 PM

The start of school is about to begin, and it’s not just the students who are about to start the new semester who should take care of themselves, but also the large AI models. Some time ago, Reddit was filled with netizens complaining that Claude was getting lazy. "Its level has dropped a lot, it often pauses, and even the output becomes very short. In the first week of release, it could translate a full 4-page document at once, but now it can't even output half a page!" https:// www.reddit.com/r/ClaudeAI/comments/1by8rw8/something_just_feels_wrong_with_claude_in_the/ in a post titled "Totally disappointed with Claude", full of

At the World Robot Conference, this domestic robot carrying 'the hope of future elderly care' was surrounded At the World Robot Conference, this domestic robot carrying 'the hope of future elderly care' was surrounded Aug 22, 2024 pm 10:35 PM

At the World Robot Conference being held in Beijing, the display of humanoid robots has become the absolute focus of the scene. At the Stardust Intelligent booth, the AI ​​robot assistant S1 performed three major performances of dulcimer, martial arts, and calligraphy in one exhibition area, capable of both literary and martial arts. , attracted a large number of professional audiences and media. The elegant playing on the elastic strings allows the S1 to demonstrate fine operation and absolute control with speed, strength and precision. CCTV News conducted a special report on the imitation learning and intelligent control behind "Calligraphy". Company founder Lai Jie explained that behind the silky movements, the hardware side pursues the best force control and the most human-like body indicators (speed, load) etc.), but on the AI ​​side, the real movement data of people is collected, allowing the robot to become stronger when it encounters a strong situation and learn to evolve quickly. And agile

ACL 2024 Awards Announced: One of the Best Papers on Oracle Deciphering by HuaTech, GloVe Time Test Award ACL 2024 Awards Announced: One of the Best Papers on Oracle Deciphering by HuaTech, GloVe Time Test Award Aug 15, 2024 pm 04:37 PM

At this ACL conference, contributors have gained a lot. The six-day ACL2024 is being held in Bangkok, Thailand. ACL is the top international conference in the field of computational linguistics and natural language processing. It is organized by the International Association for Computational Linguistics and is held annually. ACL has always ranked first in academic influence in the field of NLP, and it is also a CCF-A recommended conference. This year's ACL conference is the 62nd and has received more than 400 cutting-edge works in the field of NLP. Yesterday afternoon, the conference announced the best paper and other awards. This time, there are 7 Best Paper Awards (two unpublished), 1 Best Theme Paper Award, and 35 Outstanding Paper Awards. The conference also awarded 3 Resource Paper Awards (ResourceAward) and Social Impact Award (

Li Feifei's team proposed ReKep to give robots spatial intelligence and integrate GPT-4o Li Feifei's team proposed ReKep to give robots spatial intelligence and integrate GPT-4o Sep 03, 2024 pm 05:18 PM

Deep integration of vision and robot learning. When two robot hands work together smoothly to fold clothes, pour tea, and pack shoes, coupled with the 1X humanoid robot NEO that has been making headlines recently, you may have a feeling: we seem to be entering the age of robots. In fact, these silky movements are the product of advanced robotic technology + exquisite frame design + multi-modal large models. We know that useful robots often require complex and exquisite interactions with the environment, and the environment can be represented as constraints in the spatial and temporal domains. For example, if you want a robot to pour tea, the robot first needs to grasp the handle of the teapot and keep it upright without spilling the tea, then move it smoothly until the mouth of the pot is aligned with the mouth of the cup, and then tilt the teapot at a certain angle. . this

Hongmeng Smart Travel S9 and full-scenario new product launch conference, a number of blockbuster new products were released together Hongmeng Smart Travel S9 and full-scenario new product launch conference, a number of blockbuster new products were released together Aug 08, 2024 am 07:02 AM

This afternoon, Hongmeng Zhixing officially welcomed new brands and new cars. On August 6, Huawei held the Hongmeng Smart Xingxing S9 and Huawei full-scenario new product launch conference, bringing the panoramic smart flagship sedan Xiangjie S9, the new M7Pro and Huawei novaFlip, MatePad Pro 12.2 inches, the new MatePad Air, Huawei Bisheng With many new all-scenario smart products including the laser printer X1 series, FreeBuds6i, WATCHFIT3 and smart screen S5Pro, from smart travel, smart office to smart wear, Huawei continues to build a full-scenario smart ecosystem to bring consumers a smart experience of the Internet of Everything. Hongmeng Zhixing: In-depth empowerment to promote the upgrading of the smart car industry Huawei joins hands with Chinese automotive industry partners to provide

Distributed Artificial Intelligence Conference DAI 2024 Call for Papers: Agent Day, Richard Sutton, the father of reinforcement learning, will attend! Yan Shuicheng, Sergey Levine and DeepMind scientists will give keynote speeches Distributed Artificial Intelligence Conference DAI 2024 Call for Papers: Agent Day, Richard Sutton, the father of reinforcement learning, will attend! Yan Shuicheng, Sergey Levine and DeepMind scientists will give keynote speeches Aug 22, 2024 pm 08:02 PM

Conference Introduction With the rapid development of science and technology, artificial intelligence has become an important force in promoting social progress. In this era, we are fortunate to witness and participate in the innovation and application of Distributed Artificial Intelligence (DAI). Distributed artificial intelligence is an important branch of the field of artificial intelligence, which has attracted more and more attention in recent years. Agents based on large language models (LLM) have suddenly emerged. By combining the powerful language understanding and generation capabilities of large models, they have shown great potential in natural language interaction, knowledge reasoning, task planning, etc. AIAgent is taking over the big language model and has become a hot topic in the current AI circle. Au

See all articles