(1D) Ordered Tokens Enable Efficient Test-Time Search
This demo illustrates test-time search over 1D ordered tokens for autoregressive image generation. Ordered tokenizers such as FlexTok produce coarse-to-fine intermediate reconstructions, making partial generations meaningful to evaluate and guide with verifiers. Use this interface to compare direct AR generation with search-guided generation across different AR priors, FlexTok AR variants, search algorithms, and verifiers; the right panels show the best candidate at each search step and the search tree.
ℹ️ Performance note: This hosted demo prioritizes simplicity and speed over performance. Increasing the number of candidates per beam and the beam width under search settings gives you more control, but it also increases runtime. Search with uniform or unconditional priors may take longer because they must explore a larger candidate space. If a demo run expires due to the online time limit, please clone the Hugging Face Space and run it locally.
Text Control
Image Control (Zero Shot)
Uploading an image switches the default verifier to DreamSim. You can then add other verifiers manually.