What is t2i-l2p?

This is a research codebase from Tencent Youtu Research for a text-to-image generation method called L2P. The core idea involves taking an AI model that was originally trained to generate images by working in a compressed internal representation (called latent space) and transferring it so that it instead generates images directly in pixel space. The paper claims this transfer requires minimal extra data and computation. Most modern AI image generators work by creating images in a compressed form and then decoding that into pixels. L2P produces pixels directly. The released model is six billion parameters in size and currently generates images at 1024 by 1024 pixel resolution. The roadmap lists higher resolutions (4K, 8K, 10K) as planned future work. A live demo is hosted on Hugging Face Spaces, and the model weights are also available through Hugging Face. The repository contains code for both inference and training. To generate an image you load the model weights, point the pipeline at a text prompt, and the model produces an image. The inference code is written in Python and depends on a toolkit called DiffSynth-Studio. A Gradio web interface is also provided for running the model through a browser, with support for distributing requests across multiple GPUs automatically. Training your own version follows four steps: download a base model called Z-Image-Turbo, run a weight conversion script that adapts those latent-space weights into a pixel-space starting point, run the training script against your own image dataset, and then merge the trained output back together for inference. A low-VRAM training script is available for single GPUs with less than 24 GB of memory. This repository is primarily aimed at AI researchers and engineers working on image generation. It accompanies an arXiv preprint and is not a consumer-facing product.