The paper for Stable Diffusion 3 is finally here!

This model was released two weeks ago and uses the same DiT (Diffusion Transformer) architecture as Sora. It caused quite a stir upon release.

Compared with the previous version, the quality of images generated by Stable Diffusion 3 has been significantly improved. It now supports multi-theme prompts, and the text writing effect has also been improved, and garbled characters no longer appear. Condition.

Stability AI pointed out that Stable Diffusion 3 is a series of models with parameter sizes ranging from 800M to 8B. This parameter range means that the model can be run directly on many portable devices, significantly lowering the threshold for using large AI models.

In a newly released paper, Stability AI said that in human preference-based evaluations, Stable Diffusion 3 outperformed current state-of-the-art text-to-image generation systems such as DALL?E 3. Midjourney v6 and Ideogram v1. Soon, they will make the experimental data, code, and model weights of the study publicly available.

In the paper, Stability AI revealed more details about Stable Diffusion 3.

• ##Paper title: Scaling Rectified Flow Transformers for High-Resolution Image Synthesis
• Paper link: https://stabilityai-public-packages.s3.us-west-2.amazonaws.com/Stable Diffusion 3 Paper.pdf

Architectural details

For text-to-image generation, the Stable Diffusion 3 model must consider both text and image modes. Therefore, the authors of the paper call this new architecture MMDiT, referring to its ability to handle multiple modalities. As with previous versions of Stable Diffusion, the authors use pre-trained models to derive suitable text and image representations. Specifically, they used three different text embedding models—two CLIP models and T5—to encode text representations, and an improved autoencoding model to encode image tokens.

Stable Diffusion 3 model architecture.

Improved multimodal diffusion transformer: MMDiT block.

The SD3 architecture is based on DiT proposed by Sora core R&D member William Peebles and Xie Saining, assistant professor of computer science at New York University. Since text embedding and image embedding are conceptually very different, the authors of SD3 use two different sets of weights for the two modalities. As shown in the figure above, this is equivalent to setting up two independent transformers for each modality, but combining the sequences of the two modalities for attention operations, so that both representations can work in their own space, Another representation is also taken into account.

The author's proposed MMDiT architecture outperforms mature textual frameworks such as UViT and DiT when measuring visual fidelity and text alignment during training. to the image backbone.

In this way, information can flow between image and text tokens, thereby improving the overall understanding of the model and improving the typography of the generated output. As discussed in the paper, this architecture is also easily extensible to multiple modalities such as video.

Thanks to Stable Diffusion 3’s improved prompt following capabilities, the new model has the ability to produce images that focus on a variety of different themes and qualities, At the same time, it can also handle the style of the image itself with a high degree of flexibility.

Improve Rectified Flow through re-weighting

Stable Diffusion 3 uses the Rectified Flow (RF) formula. During the training process, Data and noise are connected in a linear trajectory. This makes the inference path straighter, thus reducing sampling steps. In addition, the authors also introduce a new trajectory sampling scheme during the training process. They hypothesized that the middle part of the trajectory would pose a more challenging prediction task, so the scheme gave more weight to the middle part of the trajectory. They compared using multiple datasets, metrics and sampler settings and tested their proposed method against 60 other diffusion trajectories such as LDM, EDM and ADM. The results show that while the performance of previous RF formulations improves with few sampling steps, their relative performance decreases as the number of steps increases. In contrast, the reweighted RF variant proposed by the authors consistently improves performance.

Extended Rectified Flow Transformer model

The author uses the reweighted Rectified Flow formula and MMDiT backbone pair Text-to-image synthesis is studied in scaling. They trained models ranging from 15 blocks with 450M parameters to 38 blocks with 8B parameters and observed that the validation loss decreased smoothly with increasing model size and training steps (first part of the figure above OK). To examine whether this translated into meaningful improvements in model output, the authors also evaluated the automatic image alignment metric (GenEval) and the human preference score (ELO) (second row above). The results show a strong correlation between these metrics and validation loss, suggesting that the latter is a good predictor of the overall performance of the model. Furthermore, the scaling trend shows no signs of saturation, making the authors optimistic about continuing to improve model performance in the future.

Flexible text encoder

By removing memory intensive 4.7B parameter T5 text encoder for inference, SD3 memory Demand can be significantly reduced with minimal performance loss. As shown, removing this text encoder has no impact on visual aesthetics (50% win rate without T5) and only slightly reduces text consistency (46% win rate). However, the authors recommend adding T5 when generating written text to fully utilize the performance of SD3, because they observed that without adding T5, the performance of generating typesetting dropped even more (win rate 38%), as shown in the following figure:

#Removing T5 for inference will only result in a significant decrease in performance when presenting very complex prompts involving many details or large amounts of written text. The image above shows three random samples of each example.

Model performance

The author compared the output image of Stable Diffusion 3 with various other open source models (including SDXL, SDXL Turbo, Stable Cascade, Playground v2.5 and Pixart-α) as well as closed-source models such as DALL-E 3, Midjourney v6 and Ideogram v1 were compared to evaluate performance based on human feedback. In these tests, human evaluators are given examples of output from each model and judged on how well the model output follows the context of the prompt given (prompt following), how well the text is rendered according to the prompt (typography), and which image Images with higher visual aesthetics are selected for the best results.

#Using SD3 as the benchmark, this chart outlines its win rate based on human evaluation of visual aesthetics, prompt following, and text layout.

From the test results, the author found that Stable Diffusion 3 is equivalent to or even better than the current state-of-the-art text-to-image generation systems in all the above aspects.

In early unoptimized inference testing on consumer hardware, the largest 8B parameter SD3 model fit the RTX 4090's 24GB VRAM, using 50 sampling steps to generate a resolution of 1024x1024 Image takes 34 seconds.

Additionally, at initial release, Stable Diffusion 3 will be available in multiple variants, ranging from 800m to 8B parametric models to further eliminate hardware barriers.

Please refer to the original paper for more details.

Reference link: https://stability.ai/news/stable-diffusion-3-research-paper