SD3
SD3（Stable Diffusion 3）是由Stability AI开发的第三代文本生成图像扩散模型。该模型采用多模态扩散变换器（MMDiT）架构，在文本理解、图像质量与细节还原方面相较前代有所提升，能够更准确地解析复杂提示词并生成高保真图像。SD3支持多种分辨率输出，适用于艺术创作、设计辅助等场景，是Stable Diffusion系列的重要迭代版本。
Timeline (last 90 days)
In MMDiTs (such as SD3 and FLUX), image tokens and text tokens are concatenated together to jointly participate in the same self-attention computation, allowing both to bidirectionally modify each other, which differs from the unidirectional cross-attention used in earlier Stable Diffusion models.
Representative DiT-class generative models include Stability AI's DiT series as well as Flux and SD3.
CFG=4–4.5 is the recommended starting range for many Flux and SD3 series models, while SDXL-type models commonly use 5–8.
Models such as FLUX and SD3 have improved text generation capabilities by introducing text encoders and more refined training data.
Stable Diffusion 1.x uses a 4-channel latent space, while next-generation models such as SD3/FLUX have upgraded to 16 channels.
Standard scales such as the SD3 (Short Dark Triad) are commonly used to measure Dark Triad traits in human psychological assessments.
The shift value of 2.8 follows the SD3/RAE rule calculated as √(32·32·32/4096), corresponding to the 32-channel latent space used by FLUX.2.
All Facts (10)
sd-scripts支持LoRA、LyCORIS（LoHa、LoKr等）、Textual Inversion、DreamBooth等微调方法，并支持SD 1.5、SDXL和SD3
90%UnverifiedSD3引入了MMDiT（Multi-Modal Diffusion Transformer）架构，用Transformer替代了传统的U-Net骨干网络，并采用三重文本编码器（CLIP ViT-L、OpenCLIP ViT-bigG、T5-XXL）
90%UnverifiedBetween 2023 and 2024, the image generation field experienced an architectural leap from Stable Diffusion 1.5 to SDXL and then to SD3.
60%UnverifiedIn MMDiTs (such as SD3 and FLUX), image tokens and text tokens are concatenated together to jointly participate in the same self-attention computation, allowing both to bidirectionally modify each other, which differs from the unidirectional cross-attention used in earlier Stable Diffusion models.
50%UnverifiedRepresentative DiT-class generative models include Stability AI's DiT series as well as Flux and SD3.
50%UnverifiedCFG=4–4.5 is the recommended starting range for many Flux and SD3 series models, while SDXL-type models commonly use 5–8.
50%UnverifiedModels such as FLUX and SD3 have improved text generation capabilities by introducing text encoders and more refined training data.
50%UnverifiedStable Diffusion 1.x uses a 4-channel latent space, while next-generation models such as SD3/FLUX have upgraded to 16 channels.
50%UnverifiedStandard scales such as the SD3 (Short Dark Triad) are commonly used to measure Dark Triad traits in human psychological assessments.
50%UnverifiedThe shift value of 2.8 follows the SD3/RAE rule calculated as √(32·32·32/4096), corresponding to the 32-channel latent space used by FLUX.2.
50%