L2P: 4K and 8K Image Generation Without a VAE

Researchers from NJU's PCaLab and Tencent YoutuResearch published L2P on May 12, 2026, a framework that converts pretrained latent diffusion models into pixel space generators by removing the VAE entirely. The result is native 4K generation and 8K zero shot extrapolation from models that were trained at standard resolutions.

The paper, arXiv 2605.12013, comes from Zhennan Chen, Junwei Zhu, Xu Chen, Jiangning Zhang, and colleagues at Nanjing University and Tencent.

What a VAE Does and Why It Creates a Resolution Ceiling

A VAE encodes an image by compressing it into a compact latent vector, then decodes that vector back to an image during inference. Standard latent diffusion models use a 4x spatial downsampling factor, meaning a 1024x1024 image becomes a 256x256 latent. At 4K (4096x4096), the latent is 1024x1024, and the decode step introduces artifacts and loses fine texture detail that was compressed out during encoding.

The information loss is not a flaw in the VAE. It is the design. Compression makes the diffusion process computationally tractable. The cost is that the VAE decode step cannot recover what was not preserved in the latent. At standard resolutions the loss is acceptable. At 4K it becomes the bottleneck.

Removing the VAE Bottleneck

Latent diffusion models compress images into a compact latent representation using a Variational Autoencoder, run the diffusion process in that compressed space, then decode back to pixels. The VAE introduces a resolution ceiling because the decode step does not scale cleanly to arbitrary resolutions, and the encode step loses information that cannot be recovered at output.

L2P discards the VAE entirely. The model processes image content directly in pixel space, working at full resolution from start to finish. There is no encode and decode round trip.

How L2P Transfers Latent Models to Pixel Space

The transfer approach is designed to reuse as much of the existing pretrained model as possible. L2P freezes the intermediate layers of the source latent diffusion model and trains only shallow transformation layers that bridge latent and pixel space representations.

Large patch tokenization handles the shift from compressed latent tokens to full resolution pixel patches. Because only the transformation layers are trained, the process runs on 8 GPUs with what the authors describe as negligible computational overhead. Critically, L2P trains entirely on synthetic images generated by the source model itself. No real training data is required.

Synthetic Self Training: Why It Works

L2P trains its transformation layers entirely on images the source model generates, with no external training data from real world sources. This works because the transformation only needs to learn the statistical relationship between the source model's latent representations and its pixel level outputs, which are already aligned by definition.

The result is a transfer method that requires no proprietary datasets and can be run by anyone with access to the source model weights. A researcher with a pretrained latent diffusion model and 8 GPUs can apply the L2P framework to produce a pixel space variant without acquiring or licensing any additional training material.

Large Patch Tokenization Explained

Processing full resolution 4K images in pixel space requires handling a very large number of tokens. Large patch tokenization groups neighboring pixels into patches before feeding them to the model, reducing the sequence length the model must process.

Patch size is a tradeoff between detail retention and computational cost. L2P's patch design allows 4K native generation to run without the memory requirements that pixel level processing at every individual position would otherwise demand. The patch grouping preserves the spatial structure that matters for image quality while keeping the computation tractable on research scale hardware.

4K Native Generation

Performance Numbers

The 97.67 Percent Speed Gain Unpacked

The source latent diffusion model at 4K requires encoding, diffusion sampling in latent space, and VAE decoding. The VAE decode step at 4K is computationally expensive because it must produce every pixel in a high resolution output. L2P eliminates that step entirely.

Single step inference at 4K is 97.67 percent faster because almost all of the eliminated time was in the decode stage. For production workflows that require iterating on generation outputs before committing to a final image, this changes the pace of the iteration loop from wait and review to near real time comparison.

8K Zero Shot: Mechanism and Limits

L2P generates 4K images natively and generalizes to 8K through extrapolation. The model scales its learned pixel space representation to a higher resolution without being trained at that resolution, because the structural relationships it learned at 4K remain valid at 8K.

The limit is that extrapolation degrades gracefully. 8K output scores below 4K on fine texture metrics but remains usable for establishing shots and large background elements where pixel level detail is less critical than overall composition and spatial coherence. For foreground subjects requiring fine texture at 8K, a dedicated high resolution model is more reliable than extrapolation from 4K.

The zero shot generalization is still a meaningful capability for establishing resolution headroom. A production that needs to deliver 8K assets for a specific distribution requirement can use L2P as a generation step and apply targeted sharpening in post, rather than rebuilding the pipeline around a model trained natively at 8K.

Access and License

L2P code is available on GitHub at TencentYoutuResearch/T2I-L2P. Weights are hosted on Hugging Face. The repository does not specify an explicit open source license, so commercial use terms should be confirmed directly with the research team before deployment.

For filmmakers using AI generated images in production, L2P represents a direct path from existing latent diffusion checkpoints to 4K and 8K capable pixel space generators. NVIDIA's PiD decoder solves the same high resolution problem from the decode end, replacing only the VAE decode step while keeping the latent generation pipeline. Microsoft's Lens approaches the efficiency question from the training side, reaching competitive results at 3.8 billion parameters on 19.3% of standard compute.

For open source pixel native generation at 8 billion parameters under MIT license, HiDream-O1-Image and the DyPE method offer additional approaches to high resolution output without VAE constraints.

NJU PCaLab and Tencent YoutuResearch

NJU PCaLab is the Perception, Cognition, and Action Lab at Nanjing University, focused on computer vision and generative modeling. Tencent YoutuResearch is the research division of Tencent's video platform. Academic and commercial research collaborations of this type have produced reproducible results in open image generation, because commercial partners provide computational resources while academic teams maintain publication standards.

The author list, Zhennan Chen, Junwei Zhu, Xu Chen, and Jiangning Zhang, spans both institutions. Publication on arXiv on May 12, 2026, with code on GitHub under the TencentYoutuResearch organization, follows the open release pattern that has made recent Chinese AI research accessible to the global open source community.

The collaboration between a university perception lab and a platform's research division is a structural advantage for this type of work. The perceptual quality questions that matter most for image generation require both the rigorous evaluation methods that academic computer vision labs apply and the large scale generation infrastructure that commercial partners operate. L2P required both to establish the GenEval and DPG-Bench results that make the claims about 4K and 8K quality verifiable.

Production Use Cases for Pixel Native 4K

Theatrical DCP masters are delivered at 4K (4096x2160). Streaming platforms deliver 4K at 3840x2160. AI generated backgrounds, environments, and VFX elements that enter these pipelines at 4K without upscaling preserve detail that upscaled latent outputs lose.

For productions using AI generated imagery in theatrical or premium streaming distribution, L2P's native 4K output means the AI generated asset meets the technical delivery requirement without an additional upscaling pass. The 97.67 percent speed advantage also means iteration on those assets fits within a normal compositing workflow rather than requiring overnight generation queues.

A production that generates 50 background plates at 4K is running 50 decode operations that the VAE would have required. At 97.67 percent faster per operation, that represents a significant reduction in GPU time across the compositing pipeline. For productions billing compute costs per hour, L2P's speed gain translates directly to budget.

To work with the latest image generation models in a browser without local setup, AI FILMS Studio provides cloud access to the image workspace.

Sources

arXiv: L2P: Unlocking Latent Potential for Pixel Generation Project Page: NJU PCaLab: L2P GitHub: TencentYoutuResearch/T2I-L2P