CogView

Why we included this project

CogView is the research code behind a 4-billion-parameter transformer that generates images from text prompts, published at NeurIPS 2021. The repo includes pretrained weights, a VQ-VAE image tokenizer, and scripts for inference, super-resolution, and image-to-text reranking, so you can run the whole pipeline instead of just reading about it. For teams studying how large autoregressive transformers behave on visual generation, the training tricks matter as much as the results: the paper describes PB-relax and Sandwich-LN, techniques that keep deep transformers stable. One honest caveat: the released model is tuned for simplified Chinese input, so it suits Chinese-language generation work or architecture study better than a drop-in English text-to-image service. The post-selection scoring step also shows a complete generation workflow worth copying when you build your own.