Collected sources and patterns will appear here. Add from search or the patterns library.
(LatentFeatures, GroundingTokens) -> GroundedLatentFeatures
Inject spatial layout constraints into a frozen diffusion model's latent representation by feeding them through an auxiliary, gated self-attention layer trained alongside the frozen layers.
Problem it solves
Fine-tuning frozen text-to-image weights directly on layout datasets causes catastrophic forgetting of general image-generation quality.
Consumes
Emits
The real projects this mechanism was found in. Attribution is the point — this is how the best teams actually do it.