nju-pcalab/l2p

L2P is a research project for generating images directly in pixel space using a diffusion model, the kind of AI model that builds an image step by step from noise. Most modern image generators work in a compressed representation called latent space to save computing power, but L2P adapts an existing pretrained model to generate pixels directly, which the authors say improves quality with only a small amount of extra data and training. The project is described in an accompanying technical report and paper. The code lets you load the pretrained model along with a text encoder and generate a 1024 by 1024 image from a text prompt, using an example script provided in the README. There is also a Gradio web demo that automatically spreads requests across whichever GPUs are free, so multiple people can use it through a browser interface at once. On the training side, there is a standard training script and a separate low VRAM version meant for a single GPU with less than 24 gigabytes of memory, and datasets are provided as a folder of images plus a CSV file listing each image's caption. The project's roadmap lists further work still to come, including releasing training code more broadly, support for much higher resolutions such as 4K, 8K, and 10K, and compatibility with more text to image diffusion models beyond the one currently supported, so some of these capabilities are not yet available in the current release. This project is aimed at AI researchers and practitioners working on image generation and diffusion models who want to experiment with pixel-space generation, reproduce the paper's results, or build on the released code and pretrained weights.