Search papers, labs, and topics across Lattice.
4
0
7
2
Current video generation models fall short of capturing the full distribution of possible behaviors, revealing a critical gap in probabilistic alignment that needs to be addressed.
DynEval reveals that a compact evaluator can achieve superior alignment with human judgments by leveraging dynamic datasets and curriculum learning strategies.
Flash-BoN reveals that spending compute on broader exploration rather than repeated verification can yield substantial performance gains in text-to-image generation.
Instruction-based image editing can now produce more physically plausible results, thanks to a new method that treats editing as a dynamic physical state transition rather than a static image mapping.
