成就了Sora和Stable diffusion 3的DiTs：究竟是啥

Miao Zheng, the author, just released it not long after, which is undoubtedly a new year for people who use artificial intelligence for creative design. So this article is specially prepared for these users to use two characteristic diffusion models and flow matching in more straightforward words to help you better use it to create diffusion models after the model is released. We will call it that as the name, and you know it is an image latent variable diffusion model based on architecture. If you have read the articles of Silicon Star, you can use big words to reveal the secrets. The method of understanding video model has realized the emergence of the physical world, so you are already at the class representative level for the following content. Just like that, it also uses the concept of block, but because it is used to generate pictures, it doesn't need to maintain the logical association between different frames of pictures like that, so it doesn't need to generate the spatio-temporal block of time and space. For me, it is similar to the one that caused a bloody storm in the field of computer vision years ago, and the images will be divided into multiple and merged. Embedded in the continuous vector space to form a sequence of input for processing, but it should be noted here that because there is business, it is necessary to receive and integrate external conditional information such as category labels or text descriptions for conditional image generation tasks, which is usually realized by providing additional input tags or cross-attention mechanisms, so that the model can guide the generation process according to the given conditional information. Then, when the block reaches the inside, it can be processed internally into the required content, which is the most important. The core part is a special structure designed for diffusion model, which can handle images and conditional information. Generally speaking, it is translated into blocks, but in order to distinguish it from others, I use generation directly and divide it into three small cross-attention. Cross-attention refers to the addition of an extra multi-head cross-attention layer after the multi-head self-attention layer. Its function is to use conditional information to guide image generation and make the generated images more consistent with prompts, but at the cost of an increase of about. The amount of calculation refers to the problem of reducing the internal covariant deviation by standardizing the output of the internal units of each layer of neural network, thus improving the convergence speed and performance in the model training process, that is, the expansion of the normalization of the standard layer, which allows the parameters of the normalization of the layer to be dynamically adjusted according to the input data or additional condition information. It is used to increase the stability and adaptability of the model, just like the car suspension. Next, an improvement is made on the basis of regression sum. In addition, the scaling parameters of dimension level are regressed and applied immediately before any residuals are connected. The purpose of this one is to imitate the beneficial initialization strategy in the residual network to promote the effective training and optimization of the model. After that, the sequence will be decoded into output noise prediction and output diagonal covariance prediction. The size of these two prediction results is the same as the spatial dimension of the input image through the standard linear decoder. Finally, these decoded tokens will be in accordance with their original. The spatial layout is rearranged to get the predicted noise value and covariance value. The second chapter is flow matching. The following abbreviation is based on the statement that it is an efficient model training method without simulation, which allows the use of general probability paths to supervise the training process. It is particularly important to break the extensible training obstacles outside the diffusion model and directly operate the probability paths without in-depth understanding of the diffusion process, thus bypassing the problems in traditional training. The so-called continuous normalized flow is a kind of deep learning. In this paper, the probability model and generation model technology of the new model transform the simple probability distribution into the probability distribution of complex high-dimensional data through a series of reversible and continuous transformations. These transformations are usually parameterized by a neural network, so that the original random variable can simulate the target data distribution after continuous transformation and be translated into vernacular to generate data like rolling dice, but it needs a lot of computing resources and time in actual operation, so I wonder if it can be another result. Almost the same as it is, but the process should be stable and the calculation amount should be low, so the essence of the method is a technology for training the model to adapt to and simulate the evolution process of a given data distribution. Even if we don't know the specific mathematical expression of this distribution or the corresponding generating vector field in advance, we can gradually let the model learn a vector field that can generate a probability distribution similar to the real data distribution by optimizing the objective function. Compared with it, it should be regarded as an optimization method. It is that the vector field generated by the training model is as close as possible to the vector field on the ideal target probability path. After reading the two core technical characteristics, you will find that it is actually very close. Both models are models, both of which used epoch-making stability and optimization, and the date of birth is so close. I don't think it is too much to say that they are related, but there is a fundamental difference between the two brothers, that is, closed source and open source. In fact, they are both closed sources. Some are open source. If you pay attention to open source artificial intelligence, then you must find that the open source community has been in trouble for a long time, and there is no obvious breakthrough. Many people have lost confidence in this and only improved the aesthetics of the generated pictures, but they can already do this. The revolutionary improvement seen can rekindle the confidence of many developers in the open source community and generate another exciting Mohammad Aymard Mostak said in Twitter that although the resources in the field of artificial intelligence are better than their own, Some of his companies have been reduced by a factor of 100, but the architecture can accept content other than video and images, but at present, I can't publish too much. I can understand what you said about pictures and videos, but what I can think of is audio. It's confusing to generate pictures through a piece of sound, but once the latest research results are released, we will definitely interpret them and generate them as soon as possible. 比特币今日价格行情网_okx交易所app_永续合约_比特币怎么买卖交易_虚拟币交易所平台

注册有任何问题请添加 微信：MVIP619 拉你进入群

文章转载注明 网址：https://netpsp.com/?id=60505