What DiT names
DiT names the denoising backbone, not the entire image pipeline. A practical system can still include text encoders, a VAE, scheduler, guidance, and optional control modules. The original DiT research used transformer blocks over latent patches and studied scaling with depth, width, and token count.
A patch projection turns regions of the latent into token-like vectors. The network also needs information about the current noise or time level and whatever condition the release supports. After the transformer predicts the denoising target, the scheduler performs the numerical update.
DiT versus U-Net
A U-Net uses multi-scale convolutional paths and skip connections. A DiT represents the latent as tokens and relies heavily on attention and feed-forward blocks. Both can perform iterative denoising; their weight shapes, memory behavior, and adapters are not compatible.
Replacing the backbone does not require every other part of diffusion to disappear. A DiT can still generate in latent space and use iterative steps. Conversely, a transformer paired with rectified flow uses a related generation loop with a different training formulation.
Modern variants
Stable Diffusion 3/3.5 uses an MMDiT design with separate text and image parameter streams that communicate. FLUX.1 is documented as a rectified-flow transformer. These are related transformer systems, not identical implementations. See the architecture dossier.
Conditioning also differs. Some DiTs use adaptive normalization, some cross-attention, and some joint or multimodal attention. A model-family name and its model card are needed to determine the actual inputs, encoders, prediction target, and scheduler.
Practical significance
Large DiT weights and attention can increase local memory pressure, making precision, quantization, and offloading important. Attention memory also grows with the number of latent patches, so resolution and batch size remain practical constraints.
A DiT label does not disclose training data, prompt safety, quality, identity consistency, or commercial rights. It also does not make LoRAs or controls portable across transformer families. Model cards and loader documentation remain necessary.