https://jalammar.github.io/illustrated-stable-diffusion/

This is a gentle introduction to how Stable Diffusion works.

text-to-image，text2img，T2I

paradise /ˈpærədaɪs/ n. 天堂，乐园 (指美好的环境)，(某些宗教所指的) 天国，乐土，伊甸园，(某类活动或某类人的) 完美去处
cosmic /ˈkɑːzmɪk/ adj. 宇宙的，巨大且重要的
beach /biːtʃ/ n. 海滩，沙滩，海滨，湖滨 v. (使) 上岸，把...拖上岸

Stable Diffusion is versatile in that it can be used in a number of different ways. Let’s focus at first on image generation from text only (text2img). The image above shows an example text input and the resulting generated image (The actual complete prompt is here). Aside from text to image, another main way of using it is by making it alter images (so inputs are text + image).
除了将文本转换为图像之外，另一种主要的使用方式是让它改变图像 (因此输入是文本 + 图像)。

versatile /ˈvɜːrsətl/ adj. 多功能的，多才多艺的，多用途的，多面手的，有多种技能的
alter /ˈɔːltər/ v. 改变，(使) 更改，修改 (衣服使更合身)，改动
pirate /ˈpaɪrət/ n. 海盗，盗版者，盗印者，道德败坏者，违法者，侵犯专利权者，非法仿制者，非法播音者 v. 当海盗，从事劫掠，盗印，窃用，以海盗方式劫掠，抢掠 adj. 盗版的，盗用的，剽窃的

1. The components of Stable Diffusion

Stable Diffusion is a system made up of several components and models. It is not one monolithic model.

monolithic /ˌmɑːnəˈlɪθɪk/ adj. 庞大而单一的，整体式的，单一的，独块巨石的，整料的，由块料组成的，单片的，单块的，庞大而无特点的，巨大而单调的 n. 单片电路，单块电路

As we look under the hood, the first observation we can make is that there’s a text-understanding component that translates the text information into a numeric representation that captures the ideas in the text.

hood /hʊd/ n. (设备或机器的) 防护罩，罩，风帽，兜帽 (外衣的一部分，可拉起蒙住头颈)，(布质) 面罩，街区，学位连领帽 (表示学位种类)，(汽车等的) 折叠式车篷 vt. 罩上，覆盖

We’re starting with a high-level view and we’ll get into more machine learning details later in this article. However, we can say that this text encoder is a special Transformer language model (technically: the text encoder of a CLIP model). It takes the input text and outputs a list of numbers representing each word/token in the text (a vector per token).

That information is then presented to the Image Generator, which is composed of a couple of components itself.

The image generator goes through two stages:

1.1. Image information creator

This component is the secret sauce of Stable Diffusion. It’s where a lot of the performance gain over previous models is achieved.

sauce /sɔːs/ n. 酱，调味汁，无礼的话 (或举动)，讨厌的话 (或举动) vt. 对...无礼，给...增加趣味或风味，调味或加沙司于...

This component runs for multiple steps to generate image information. This is the steps parameter in Stable Diffusion interfaces and libraries which often defaults to 50 or 100.

The image information creator works completely in the image information space (or latent space). We’ll talk more about what that means later in the post. This property makes it faster than previous diffusion models that worked in pixel space. In technical terms, this component is made up of a UNet neural network and a scheduling algorithm.

The word “diffusion” describes what happens in this component. It is the step by step processing of information that leads to a high-quality image being generated in the end (by the next component, the image decoder).

1.2. Image Decoder

The image decoder paints a picture from the information it got from the information creator. It runs only once at the end of the process to produce the final pixel image.

With this we come to see the three main components (each with its own neural network) that make up Stable Diffusion:

• ClipText for text encoding

Input: text.
Output: 77 token embeddings vectors, each in 768 dimensions.

• UNet + Scheduler to gradually process/diffuse information in the information (latent) space

Input: text embeddings and a starting multi-dimensional array (structured lists of numbers, also called a tensor) made up of noise.
Output: A processed information array

• Autoencoder Decoder that paints the final image using the processed information array

Input: The processed information array (dimensions: (4, 64, 64))
Output: The resulting image (dimensions: (3, 512, 512) which are (red/green/blue, width, height))

diffuse /dɪˈfjuːs , dɪˈfjuːz/ adj. 扩散的，漫射的，弥漫的，不清楚的，冗长的，难解的，啰唆的 v. (使气体或液体) 扩散，弥漫，渗透，(使光) 模糊，漫射，漫散，传播，使分散，散布，普及

2. What is Diffusion anyway?

Diffusion is the process that takes place inside the pink “image information creator” component. Having the token embeddings that represent the input text, and a random starting image information array (these are also called latents), the process produces an information array that the image decoder uses to paint the final image.

This process happens in a step-by-step fashion. Each step adds more relevant information. To get an intuition of the process, we can inspect the random latents array, and see that it translates to visual noise. Visual inspection in this case is passing it through the image decoder.

intuition /ˌɪntuˈɪʃn/ n. (一种) 直觉，直觉力

Diffusion happens in multiple steps, each step operates on an input latents array, and produces another latents array that better resembles the input text and all the visual information the model picked up from all images the model was trained on.

resemble /rɪˈzembl/ vt. 像，类似于，看起来像，显得像

We can visualize a set of these latents to see what information gets added at each step.

The process is quite breathtaking to look at.

Something especially fascinating happens between steps 2 and 4 in this case. It’s as if the outline emerges from the noise.

emerge /iˈmɜːrdʒ/ v. (从隐蔽处或暗处) 出现，浮现，显现，暴露，露出，显露，被知晓，幸存下来，摆脱出来，露头，露出真相

2.1. How does Diffusion work?

The central idea of generating images with diffusion models relies on the fact that we have powerful computer vision models. Given a large enough dataset, these models can learn complex operations. Diffusion models approach image generation by framing the problem as following:

Say we have an image, we generate some noise, and add it to the image.

This can now be considered a training example. We can use this same formula to create lots of training examples to train the central component of our image generation model.

While this example shows a few noise amount values from image (amount 0, no noise) to total noise (amount 4, total noise), we can easily control how much noise to add to the image, and so we can spread it over tens of steps, creating tens of training examples per image for all the images in a training dataset.

With this dataset, we can train the noise predictor and end up with a great noise predictor that actually creates images when run in a certain configuration. A training step should look familiar if you’ve had ML exposure:

2.2. Painting images by removing noise

Let’s now see how this can generate images.

The trained noise predictor can take a noisy image, and the number of the denoising step, and is able to predict a slice of noise.

The sampled noise is predicted so that if we subtract it from the image, we get an image that’s closer to the images the model was trained on (not the exact images themselves, but the distribution - the world of pixel arrangements where the sky is usually blue and above the ground, people have two eyes, cats look a certain way - pointy ears and clearly unimpressed).

unimpressed /ˌʌnɪmˈprest/ adj. 印象平平的，无深刻印象的

If the training dataset was of aesthetically pleasing images (e.g., LAION Aesthetics https://laion.ai/blog/laion-aesthetics/, which Stable Diffusion was trained on), then the resulting image would tend to be aesthetically pleasing. If the we train it on images of logos, we end up with a logo-generating model.

aesthetical [i:s'θetɪkəl] adj. 美的，美学的，审美的
please [pliz] int. 请务必，请问，太感谢了，收敛点儿 v. 喜欢，使满意，使愉快

This concludes the description of image generation by diffusion models mostly as described in Denoising Diffusion Probabilistic Models (https://arxiv.org/abs/2006.11239). Now that you have this intuition of diffusion, you know the main components of not only Stable Diffusion, but also Dall-E 2 and Google’s Imagen.

Note that the diffusion process we described so far generates images without using any text data. So if we deploy this model, it would generate great looking images, but we’d have no way of controlling if it’s an image of a pyramid or a cat or anything else. In the next sections we’ll describe how text is incorporated in the process in order to control what type of image the model generates.
请注意，我们到目前为止描述的扩散过程无需使用任何文本数据即可生成图像。因此，如果我们部署此模型，它将生成外观精美的图像，但我们无法控制它是金字塔、猫还是其他图像。在下一节中，我们将描述如何在该过程中合并文本以控制模型生成的图像类型。

3. Speed Boost: Diffusion on compressed (latent) data instead of the pixel image

To speed up the image generation process, the Stable Diffusion paper runs the diffusion process not on the pixel images themselves, but on a compressed version of the image. The paper calls this “Departure to Latent Space” (High-Resolution Image Synthesis with Latent Diffusion Models).

This compression (and later decompression/painting) is done via an autoencoder. The autoencoder compresses the image into the latent space using its encoder, then reconstructs it using only the compressed information using the decoder.

departure [dɪˈpɑrtʃər] n. 离开，出发，背离，起程

Now the forward diffusion process is done on the compressed latents. The slices of noise are of noise applied to those latents, not to the pixel image. And so the noise predictor is actually trained to predict noise in the compressed representation (the latent space).

The forward process (using the autoencoder’s encoder) is how we generate the data to train the noise predictor. Once it’s trained, we can generate images by running the reverse process (using the autoencoder’s decoder).

These two flows are what’s shown in Figure 3 of the LDM/Stable Diffusion paper:

This figure additionally shows the “conditioning” components, which in this case is the text prompts describing what image the model should generate. So let’s dig into the text components.

4. The Text Encoder: A Transformer language model

A Transformer language model is used as the language understanding component that takes the text prompt and produces token embeddings. The released Stable Diffusion model uses ClipText (A GPT-based model), while the paper used BERT.

The Illustrated GPT-2 (Visualizing Transformer Language Models)
https://jalammar.github.io/illustrated-gpt2/

The Illustrated BERT, ELMo, and co. (How NLP Cracked Transfer Learning)
https://jalammar.github.io/illustrated-bert/

The choice of language model is shown by the Imagen paper to be an important one. Swapping in larger language models had more of an effect on generated image quality than larger image generation components.

Larger / better language models have a significant effect on the quality of image generation models. https://arxiv.org/abs/2205.11487

The early Stable Diffusion models just plugged in the pre-trained ClipText model released by OpenAI. It’s possible that future models may switch to the newly released and much larger OpenCLIP variants of CLIP (True enough, Stable Diffusion V2 uses OpenClip). This new batch includes text models of sizes up to 354M parameters, as opposed to the 63M parameters in ClipText.

plug [plʌɡ]：n. 插头，(电源) 插座，转换插头，塞子 v. 堵塞，封堵，补足，补充

LARGE SCALE OPENCLIP: L/14, H/14 AND G/14 TRAINED ON LAION-2B
https://laion.ai/blog/large-openclip/

Stable Diffusion 2.0 Release
https://stability.ai/news/stable-diffusion-v2-release

4.1. How is CLIP trained?

CLIP is trained on a dataset of images and their captions. Think of a dataset looking like this, only with 400 million images and their captions:

In actuality, CLIP was trained on images crawled from the web along with their “alt” tags.

CLIP is a combination of an image encoder and a text encoder. Its training process can be simplified to thinking of taking an image and its caption. We encode them both with the image and text encoders respectively.

actuality [ˌæktʃuˈæləti] n. 实际，真实，真实情况，现实情况
crawl [krɔl] v. 爬，爬行，匍匐行进，(昆虫) 爬行 n. 自由泳，爬泳，缓慢的速度

We then compare the resulting embeddings using cosine similarity. When we begin the training process, the similarity will be low, even if the text describes the image correctly.

We update the two models so that the next time we embed them, the resulting embeddings are similar.

By repeating this across the dataset and with large batch sizes, we end up with the encoders being able to produce embeddings where an image of a dog and the sentence “a picture of a dog” are similar. Just like in word2vec, the training process also needs to include negative examples of images and captions that don’t match, and the model needs to assign them low similarity scores.

The Illustrated Word2vec
https://jalammar.github.io/illustrated-word2vec/

5. Feeding text information into the image generation process

To make text a part of the image generation process, we have to adjust our noise predictor to use the text as an input.

Our dataset now includes the encoded text. Since we’re operating in the latent space, both the input images and predicted noise are in the latent space.

To get a better sense of how the text tokens are used in the Unet, let’s look deeper inside the Unet.

5.1. Layers of the Unet Noise predictor without text

Let’s first look at a diffusion Unet that does not use text. Its inputs and outputs would look like this:

Inside, we see that:

• The Unet is a series of layers that work on transforming the latents array
• Each layer operates on the output of the previous layer
• Some of the outputs are fed (via residual connections) into the processing later in the network
• The timestep is transformed into a time step embedding vector, and that’s what gets used in the layers

5.2. Layers of the Unet Noise predictor with text

Let’s now look how to alter this system to include attention to the text.

The main change to the system we need to add support for text inputs (technical term: text conditioning) is to add an attention layer between the ResNet blocks.

Note that the ResNet block doesn’t directly look at the text. But the attention layers merge those text representations in the latents. And now the next ResNet can utilize that incorporated text information in its processing.

6. Conclusion

I hope this gives you a good first intuition about how Stable Diffusion works. Lots of other concepts are involved, but I believe they’re easier to understand once you’re familiar with the building blocks above.

7. Resources

DreamStudio, https://beta.dreamstudio.ai/generate
The Annotated Diffusion Model, https://huggingface.co/blog/annotated-diffusion
What are Diffusion Models? https://lilianweng.github.io/posts/2021-07-11-diffusion-models/
Stable Diffusion with 🧨 Diffusers, https://huggingface.co/blog/stable_diffusion

Acknowledgements

Citation

https://jalammar.github.io/illustrated-stable-diffusion/

References

[1] Yongqiang Cheng, https://yongqiang.blog.csdn.net/