Generative Vision Atlas

Conditioning & control option

Frozen large-language-model text encoder (T5-style)

Replace or supplement the CLIP text tower with a much larger text-only pretrained language model (T5) as the source of prompt embeddings, on the finding that scaling the text encoder improves fidelity and prompt alignment more than scaling the image diffusion model does.

Mechanism

Imagen swaps a frozen, text-only pretrained language model (T5) in for a CLIP-style text tower as the source of prompt embeddings for diffusion conditioning. The paper’s central, explicitly stated finding: “generic large language models (e.g. T5), pretrained on text-only corpora, are surprisingly effective at encoding text for image synthesis: increasing the size of the language model in Imagen boosts both sample fidelity and image-text alignment much more than increasing the size of the image diffusion model.” Imagen predates this atlas’s 2023-2026 scope and is a pixel-space, not latent, diffusion model — kept as a full entity per DECISIONS.md D018 specifically for this conditioning-axis finding, which every later text-encoder choice in this atlas responds to.

Trade-offs

Parameter cost: T5-XXL is far larger than a CLIP text tower, and this cost is paid at every conditioning pass, independent of the image model’s own size. Flexibility: a text-only language model captures compositional and syntactic structure a contrastive image-text objective does not optimize for directly, which is the gap Imagen’s finding attributes the fidelity/alignment gain to. What it does not solve: T5 has no visual grounding at all — it was never trained to align with image content the way CLIP’s text tower was — which is part of why SD3 concatenates T5-XXL with CLIP embeddings rather than replacing CLIP outright, and why later systems moved to a vision-language model that has both language competence and visual grounding at once.

Introduced by

Used by (1)

Imagen

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, Joint (shared) attention conditioning, Native autoregressive token conditioning, Learnable query-token bridge, In-context sequence concatenation, Vision-language model as text encoder

← All Conditioning & control options