L2P ditches VAEs to generate images directly in pixel space

If you have spent any time generating AI images lately, you have probably accepted that tiny details and rendered text often look a bit mushy. The culprit has long been the VAE, the compression tool that unpacks math from a latent space into actual pixels.

Researchers just dropped L2P, short for Latent-to-Pixel. It takes an axe to that entire pipeline by ditching the VAE and generating images directly in pixel space.

Swapping engines with just eight GPUs

Traditional latent diffusion models do their heavy lifting inside a compressed sandbox to save compute. It keeps generation fast, but passing through a VAE bottleneck inherently strips away texture and fidelity.

Training a pure pixel model from scratch usually requires millions of dollars in compute. L2P gets around this with a surprisingly clever shortcut.

Instead of starting over, the team took pre-trained weights from existing powerhouses like Z-Image, froze the deeper layers, and trained only the shallow layers using large-patch tokenization. Using nothing but synthetic images generated by the model itself, they pulled off the entire migration on just eight GPUs with zero real-world data collection.

Native 4K output meets a 20GB checkpoint

Removing the VAE does not just fix blurry text; it removes a massive memory choke point. L2P can generate native 4K visuals and even extrapolate up to 8K resolution.

Benchmark-wise, it hits parity with source models on DPG-Bench and matches 93% performance on GenEval. In side-by-side tests, the team claims it punches above popular open-source latent engines like Qwen and Z-Image Turbo when handling complex realism and typography.

The team open-sourced the training scripts and code alongside a 1K resolution checkpoint. Keep in mind that this file clocks in at roughly 20GB, meaning you will need a beefy mid-to-high-end GPU to run it locally. Native 4K weights are still queued up for later.

Why it matters

We think L2P is a huge conceptual win. It proves that moving from latent shortcuts to pixel-perfect fidelity does not require massive compute clusters if you know how to reuse what already works.

For creators and designers, this opens the door to ditching blurred text and muddy skin textures once and for all. That said, a 20GB footprint just for 1K outputs means this is not a lightweight toy for average laptops right now.

For the broader AI ecosystem, this technique is a blueprint. If smaller labs can strip the VAE bottleneck off existing foundation models using just a single server node, the race for ultra-high-resolution generation is about to get a lot more competitive.