Multi-Resolution Tokenization: Joint 1D and 2D Token Streams for Robust Generation

TL;DR: Can we combine the strengths of 1D ordered and 2D grid tokenizations by generating both streams in parallel, using cross-stream attention and mutual verification? The concrete experiment: generate coarse 1D tokens and fine 2D tokens simultaneously, and let their intermediate states condition each other during search.

Research Question: Does fusing 1D and 2D token streams in a multi-resolution, cross-attentive autoregressive model improve the controllability and quality of image generation during test-time search?

Hypothesis: Jointly modeling and verifying both global (1D) and local (2D) features during generation will yield more robust and semantically aligned outputs than using either stream alone.

Experiment Plan: Extend the AR model to generate two parallel token streams: 1D coarse-to-fine (global) and 2D grid (local). Use cross-attention layers to allow each stream to condition on the other’s intermediate representations. Interleave verifier feedback: global verifier for 1D tokens, local verifier for 2D tokens, with cross-checks between them. Benchmark against single-stream baselines on text-to-image tasks, measuring alignment, detail, and efficiency.

References:

• Gao, Z., Rezaei, P., Cy, A., Ye, M., Jovanovi'c, N., Allardice, J., Dehghan, A., Zamir, A., Bachmann, R., & Kar, O. F. (2026). (1D) Ordered Tokens Enable Efficient Test-Time Search.