Flexible Resolution Generation With Dynamic Padding

text-to-image model by undefined. 7,16,659 downloads.

Unique: Uses position embeddings that generalize across resolutions, enabling variable-size generation without model retraining. Implements efficient dynamic padding to avoid wasted computation on non-square images.

vs others: More flexible than fixed-resolution models; comparable to other variable-resolution diffusion models but with better optimization for consumer hardware.

via “variable-resolution image processing with dynamic padding”

* ⭐ 04/2022: [Hierarchical Text-Conditional Image Generation with CLIP Latents (DALL-E 2)](https://arxiv.org/abs/2204.06125)

Unique: Implements dynamic padding that adapts to input dimensions while maintaining alignment with hierarchical window and patch structures, enabling efficient variable-resolution processing without fixed input constraints

vs others: More flexible than fixed-resolution models and more efficient than naive resizing approaches, enabling batch processing of mixed-resolution images while preserving aspect ratios