Tencent Youtu Lab's L2P method turns latent diffusion models into pixel-space generators

Tencent Youtu Lab, in collaboration with Nanjing University, has introduced L2P, a novel method that transforms existing latent diffusion models—such as Alibaba's Z-Image—into pixel-space generators. The key innovation lies in freezing the intermediate layers of the source latent diffusion model and training only its shallow layers to learn the transition from latent to pixel behavior. This approach eliminates the need for a Variational Autoencoder (VAE) decoder, which previously acted as a bottleneck in latent diffusion pipelines. By removing the VAE, L2P achieves higher-resolution outputs and significantly reduces training costs, as the majority of the model parameters remain untouched. The technique is practical for real-world deployment since it leverages pre-trained latent weights and only requires lightweight fine-tuning.

The L2P paper, published on arXiv on May 12, 2026, demonstrates that the method can generate images at resolutions beyond typical latent model capabilities without additional computational overhead. This is particularly valuable for applications demanding fine-grained detail, such as medical imaging, high-fidelity content creation, and scientific visualization. Researchers note that L2P retains the generative quality of the original latent diffusion model while offering the flexibility of pixel-space production. By bridging the gap between latent efficiency and pixel precision, L2P opens new avenues for cost-effective, high-resolution image synthesis. The approach also hints at future directions where hybrid models could dynamically switch between latent and pixel spaces depending on task requirements.