苹果，AI和图像生成：新的流状态？

苹果对使用标准化流量的AI图像生成的创新方法可以重塑设备AI的未来。

The world of AI image generation is heating up, and Apple is throwing its hat into the ring with a fresh take on an old technique. While diffusion models and autoregressive models dominate the scene, Apple's recent research suggests that Normalizing Flows, with a Transformer twist, might be the dark horse we didn't see coming. Could this be the key to powerful, on-device AI image generation?

AI图像的世界正在加热，Apple用新的旧技术将其帽子扔进了戒指。尽管扩散模型和自回旋模型占据了现场的主导，但苹果的最新研究表明，通过变压器扭曲的流动可能是我们没有看到的黑马。这可能是强大的，设备AI图像生成的关键吗？

Normalizing Flows: A Blast from the Past?

归一化流：过去的爆炸？

Normalizing Flows (NFs) aren't exactly new, but Apple's revisiting them with a modern spin. Think of NFs as AI that learns to transform images into structured noise, and then reverses the process to create new images. The cool part? They can calculate the exact likelihood of each image, something diffusion models struggle with. This is super useful where understanding probability matters.

标准化流量（NFS）并不是什么新鲜事物，但是苹果正在用现代旋转来重新审视它们。将NFS视为AI，学会将图像转换为结构化噪声，然后逆转创建新图像的过程。很酷的部分？他们可以计算每个图像的确切可能性，扩散模型与之困扰。在理解概率很重要的情况下，这非常有用。

So, why aren't NFs all the rage? Early versions produced blurry, underwhelming images. But Apple believes they've cracked the code.

那么，为什么NFS并不是所有的愤怒呢？早期版本产生了模糊，不足的图像。但是苹果认为他们已经破解了代码。

TarFlow: Transformers to the Rescue

TARFLOW：救援的变压器

Apple's first study introduces TarFlow (Transformer AutoRegressive Flow). The secret sauce? Swapping out old-school layers for Transformer blocks. TarFlow chops images into patches and generates them block by block, predicting each one based on what came before. Sound familiar? It's similar to how OpenAI generates images, but with a crucial difference.

苹果的第一项研究引入了TARFLOW（变压器自动回归流）。秘密调味料？将老式层交换为变压器块。 TARFLOW将图像置于补丁中，并通过块生成它们，并根据以前的内容预测每个图像。听起来很熟悉吗？这类似于OpenAI生成图像的方式，但差异至关重要。

Instead of generating discrete tokens (like words), TarFlow generates pixel values directly. This avoids the quality loss that comes with compressing images into a fixed vocabulary. It's a subtle but significant advantage.

Tarflow无需生成离散令牌（如单词），而是直接生成像素值。这避免了将图像压缩成固定词汇的质量损失。这是一个微妙但很大的优势。

STARFlow: Scaling Up the Vision

Starflow：扩大视觉

But TarFlow had its limits, especially with high-res images. That's where STARFlow (Scalable Transformer AutoRegressive Flow) comes in. The big change? STARFlow works on a compressed version of the image, then uses a decoder to upscale it. This "latent space" approach lets STARFlow focus on the big picture, leaving the fine details to the decoder.

但是TARFLOW具有其限制，尤其是使用高分辨率图像。那就是Starflow（可扩展的变压器自回旋流程）的出现。大变化？ Starflow在图像的压缩版本上工作，然后使用解码器对其进行更大的幅度。这种“潜在空间”方法使Starflow专注于大局，将细节留给解码器。

Apple also cleverly integrates existing language models (like Google's Gemma) for text prompts. This keeps the image generation focused on visual refinement.

苹果还巧妙地将现有语言模型（例如Google的Gemma）集成为文本提示。这使图像产生集中在视觉上。

Apple vs. OpenAI: A Different Approach

苹果与Openai：另一种方法

While Apple's rethinking flows, OpenAI's GPT-4o is also moving beyond diffusion. But their approaches are worlds apart. GPT-4o treats images as sequences of tokens, giving it incredible flexibility. The same model can generate text, images, and audio.

苹果重新思考流动时，OpenAI的GPT-4O也超越了扩散。但是他们的方法与众不同。 GPT-4O将图像视为令牌序列，从而使其具有令人难以置信的灵活性。相同的模型可以生成文本，图像和音频。

The downside? Token-by-token generation can be slow and computationally expensive. But since GPT-4o runs in the cloud, OpenAI isn't as worried about latency or power.

缺点？逐个代币的生成可能很慢，计算上的昂贵。但是，由于GPT-4O在云中运行，因此Openai不必担心潜伏期或力量。

The key takeaway: OpenAI is building for data centers, while Apple is clearly building for our iPhones. The ability to run powerful AI image generation on-device, without relying on cloud connectivity, is a game-changer. It opens up possibilities for real-time creativity and enhanced privacy.

关键要点：OpenAI正在为数据中心建造，而Apple显然正在为我们的iPhone建造。在不依赖云连接的情况下运行强大的AI图像生成功能的能力是改变游戏规则的能力。它为实时创造力和增强隐私性开辟了可能性。

Why Normalizing Flows Matter

为什么正常流量很重要

Normalizing flows offer the ability to calculate the exact likelihood of the generated image. This is a key advantage over other methods and makes flows appealing for tasks where understanding the probability of an outcome really matters.

归一化流提供了计算生成图像的确切可能性的能力。这是比其他方法的关键优势，并且使流量吸引了理解结果的可能性确实很重要的任务。

The Future of AI Image Generation on Apple Devices

AI图像在Apple设备上的未来

Apple's approach could lead to faster, more private, and energy-efficient AI image generation on its devices. Imagine creating stunning visuals on your iPhone without needing a constant internet connection or worrying about your data being sent to the cloud. That's the promise of Apple's Normalizing Flows.

苹果的方法可能会导致其设备上更快，更私密和节能的AI图像生成。想象一下，在iPhone上创建令人惊叹的视觉效果，而无需持续的Internet连接或担心将数据发送到云。这就是苹果正常化流动的承诺。

While Large Reasoning Models (LRMs) reasoning capabilities are debated, as highlighted by Apple's "Illusion of Thinking" paper and Anthropic's rebuttal, the focus on efficient, on-device AI suggests Apple is prioritizing practicality and user experience.

虽然辩论了大型推理模型（LRMS）推理能力，但苹果的“思考”纸张和人类的反驳强调了，但专注于高效的，on device AI的重点表明，苹果正在优先考虑实用性和用户体验和用户体验。

A Personal Take

个人看法

It's easy to get caught up in the hype around massive cloud-based AI models. But Apple's focus on bringing AI to our pockets is incredibly exciting. The potential for on-device AI to revolutionize creativity, productivity, and privacy is immense. I think Apple's bet on Normalizing Flows is a smart one, and I can't wait to see what they come up with next.

很容易围绕大型基于云的AI模型吞噬炒作。但是，苹果专注于将AI带到我们的口袋里的关注非常令人兴奋。在设备AI革新创造力，生产力和隐私的潜力是巨大的。我认为苹果对归一化流量的赌注是一个聪明的人，我迫不及待地想看看他们接下来想出的内容。

So, there you have it. Apple's taking a different path in the AI image generation race, and it's definitely one to watch. Who knows, maybe Normalizing Flows will be the next big thing. One thing's for sure: the future of AI is looking pretty flow-tastic!

所以，你有。苹果在AI图像生成竞赛中采取了不同的道路，这绝对是值得关注的。谁知道，也许将流量标准化将是下一个大事。可以肯定的是：AI的未来看起来很流畅！