SD 3 and 3.5
Stable Diffusion 3 and 3.5 by Stability AI: image models with three text encoders and a 16-channel latent.
What it is
SD3 Medium (2024) and SD 3.5 Large, Large Turbo and Medium (late 2024) are transformer models ("MMDiT") that read sentences well thanks to a T5 encoder next to two CLIP encoders. They are used less than SDXL and Flux, so fewer add-ons exist.
Files and nodes
Either a checkpoint that includes the text encoders (Load Checkpoint), or the model alone with TripleCLIPLoader (clip_l, clip_g, t5xxl) and its VAE. EmptySD3LatentImage (16 channels). ModelSamplingSD3 sets the shift.
Settings that work
About one megapixel (1024×1024); 28 to 40 steps; cfg about 4 to 5.5 (Large Turbo: about 4 steps, cfg near 1); euler or dpmpp_2m with sgm_uniform or simple, per the template. Prompt: sentences.
What fits
Only SD3 or SD 3.5 LoRAs and ControlNets, and only for the same size (Medium and Large differ). SDXL and Flux add-ons do not work.
How the app helps
Shown as SD 3 on model cards. The lesson SD 3.5: one file and a new empty latent builds its graph.