Generative Vision Atlas

Conditioning & control option

Joint (shared) attention conditioning

Concatenate condition and generation tokens into one sequence and let ordinary self-attention mix them, instead of routing conditioning through separate cross-attention layers.

Mechanism

Text and image tokens are concatenated into a single sequence before self-attention runs, so every token attends to every other token — condition and target are no longer architecturally distinguished once they reach the attention operation. SD3 introduces this as MM-DiT: “since text and image embeddings are conceptually quite different, we use two separate sets of weights for the two modalities… equivalent to having two independent transformers for each modality, but joining the sequences of the two modalities for the attention operation.” Lumina-Image 2.0 later shows the two-weight-set part is not required — its single-stream variant runs the same joint-sequence idea through one shared set of weights.

Trade-offs

Parameter cost: MM-DiT’s dual-stream form roughly doubles the parameters devoted to the combined forward pass relative to a single-stream design, since text and image each get their own full transformer capacity; compute cost also grows with (text+image) tokens squared rather than image tokens squared plus a linear cross-attention term. Flexibility: conditioning and generation share the same representational space by construction, which is exactly what made concatenating an entire reference image’s tokens (sequence-concat) a natural next step once this attention pattern was established. Quality: SD3’s own ablation (Sec. 4) reports this design “significantly outperforms the cross-attention and vanilla variants” it was tested against — the paper’s own head-to-head comparison, not an external benchmark.

Introduced by

Used by (4)

Lumina-Image 2.0, Qwen-Image, SD3 / MMDiT, SVG-T2I

Alternatives on this axis

adaLN / adaLN-Zero modulation, Adapter-based spatial conditioning, CLIP text encoder as conditioning source, Cross-attention conditioning, Decoupled cross-attention (image-prompt adapter), Identity-preserving face conditioning, Native autoregressive token conditioning, Learnable query-token bridge, In-context sequence concatenation, Frozen large-language-model text encoder (T5-style), Vision-language model as text encoder

← All Conditioning & control options