Variable-Length Tokenization via Learnable Global Merging for Diffusion Transformers

The article presents a novel variable-length tokenizer for Latent Diffusion Models (LDMs) that uses learnable global merging to improve token representation across varying lengths. This approach allows for adaptive compression without the semantic inconsistencies associated with traditional token truncation methods. The proposed tokenizer achieves a superior generative performance on ImageNet 256×256, demonstrating better quality-compute trade-offs than existing variable-length tokenization methods, which is critical for practitioners aiming to optimize LDMs in visual synthesis tasks.