Generative Vision Atlas

Design axis

Architecture

What backbone turns the objective into a trainable network?

DiT with diffusion head (DiT^DH)

A DiT variant whose final block is replaced by a small per-token diffusion head, used to decode high-dimensional semantic latents where a single linear projection is insufficient.

Diffusion Transformer (DiT)

A plain Vision-Transformer backbone for diffusion/flow models, operating on a grid of latent patch tokens with adaptive layer norm for timestep/class conditioning.

MM-DiT (dual-stream joint attention)

Separate weights for text and image token streams, joined only through shared self-attention layers, so each modality keeps its own representation while still attending jointly.

Single-stream (unified-weight) DiT

Give text and image tokens one shared set of transformer weights (rather than MM-DiT's two parallel weight sets) and rely purely on the joint sequence concatenation to let self-attention mix modalities, trading modality-specific capacity for a smaller, simpler backbone.

UNet backbone

A convolutional encoder-decoder with skip connections; the dominant diffusion backbone before diffusion transformers.