Family and lineage
Stable Diffusion 3 introduced Stability AI’s Multimodal Diffusion Transformer architecture. Stable Diffusion 3.5 uses the same broad MMDiT family with revised training and documented variants rather than returning to the U-Net used by SDXL. This is an official architectural succession, but adapters and checkpoints from SD 1.x or SDXL are not automatically compatible.
The SD3 paper describes separate parameter sets for text and image representations with bidirectional information exchange. It couples this architecture with rectified-flow training and reports scaling experiments across model sizes.
Conditioning and architecture
The official 3.5 Large model card identifies three fixed pretrained text encoders: CLIP-L, OpenCLIP-G, and T5-XXL. Their embeddings condition an MMDiT denoiser operating in image latent space. The card also documents query-key normalization for training stability. These facts apply to the named release; they should not be projected onto every variant without its own card.
Compared with SDXL’s convolution-heavy U-Net, the transformer backbone organizes computation around token-like latent patches and attention. That can improve scaling and text-image interaction, while the large denoiser and T5 encoder make weight loading and memory planning important for local use.
Variants, access, and local inference
The 3.5 family includes named Large, Large Turbo, and Medium variants with different size and inference characteristics. “Turbo” denotes a distilled, low-step release rather than a separate base architecture. The official card recommends ComfyUI or Diffusers for self-hosted inference and documents quantization examples.
Weights are open for inspection and local use under Stability AI’s current Community License, not an unrestricted public-domain grant. Users need to read the exact model license and any revenue thresholds. Hosted endpoints add their own retention and content policies.
Practical significance and limits
The family matters to adult-AI workflows because it offers stronger prompt/text handling than earlier Stability families and can be run locally when hardware and license terms fit. Its ecosystem is smaller and less cross-compatible than mature SDXL/SD 1.5 collections. Large text encoders and transformer weights increase download, RAM, and VRAM pressure, so quantization and offloading may trade memory for latency or output differences.
Stability AI discloses only broad training-data categories in the 3.5 card. Exact dataset composition and many training details are not public and are not inferred here. The model can still produce anatomy, typography, bias, and prompt-following failures.