chunyu0208/lpd
Accelerate autoregressive image generation with Locality-aware Parallel Decoding (LPD). Explore our code and models on GitHub! 🚀🌟
What's novel
Accelerate autoregressive image generation with Locality-aware Parallel Decoding (LPD). Explore our code and models on GitHub! 🚀🌟
Code Analysis
12 files read · 4 roundsImplements Locality-aware Parallel Decoding (LPD) for autoregressive image generation, using a spatially-aware token ordering algorithm and block-causal attention masks to generate VQGAN tokens in parallel groups rather than one-at-a-time, significantly reducing inference steps while maintaining gen
Strengths
The core algorithm (locality-aware ordering with proximity scoring + farthest-point sampling, combined with block-causal attention via flex_attention) is a genuinely novel and well-formulated approach to parallelizing autoregressive image decoding. The codebase is well-organized with clean separation between the model, ordering algorithm, training loop, and evaluation, and includes proper KV-cache inference, distributed training, and mixed precision support.
Weaknesses
Zero test coverage and a severely broken README that references a zip file URL as a local path in all usage examples, making it nearly useless for understanding the project. The ordering algorithm has O(n²) complexity per step, uses eval() for string parsing, and the farthest_point_sampling function has a subtle indexing bug.
Score Breakdown
Signal breakdown
Innovation
Craft
Traction
Scope
Evidence
Commits
5
Contributors
2
Files
41
Active weeks
3
Repository
Language
Python
Stars
1
Forks
0
License
MIT