Pixel-space text-to-image diffusion model transferred from latent diffusion models, without a VAE, supporting 4K output.
Updated EU status checked Last shown
Sources: the vendor's own pages (linked above) and creator videos at the linked timestamps. EU status is information, not legal advice; see how we check.
L2P (Latent-to-Pixel) is a research transfer method from NJU PCALab and collaborators that converts a pre-trained latent diffusion model (LDM) into a pixel-space diffusion model. It discards the VAE in favour of large-patch tokenization, freezes the source model's intermediate layers, and trains only the shallow layers to learn the latent-to-pixel mapping. Training uses only images synthesised by the source LDM, so no real-data collection is needed, and the authors report that 8 GPUs are sufficient.
According to the project page, L2P performs on par with the source LDM on DPG-Bench and reaches 93% of its performance on GenEval. Removing the VAE memory bottleneck allows native 4K generation, with a stated 97.67% faster single-step inference than the source LDM at 4K, and zero-shot extrapolation to 8K. A YouTube review describes a version built from Z-Image; it says only the 1K model (about 20 GB) had been released, with training scripts.
The material provided includes only the project page; no GitHub, Hugging Face or paper links were supplied, and the paper is cited as arXiv:2605.12013. The license and the exact release status of weights could not be confirmed from the provided text.
Watch the moment: the exact second top creators show it. Videos load only when you press play, via youtube-nocookie.com.
Pixel-space diffusion image model built from Z-Image by removing the VAE and latent space. Handles up to 4K and 8K extrapolation; only the 1K model (~20 GB) is released so far, with training scripts.
Closest neighbours by meaning, not just by category.
Model · Princeton
Fully open 3B text-to-image diffusion model from Princeton with released weights, training code, and data pipeline.
Model · Google
Google's open-weights 26B diffusion language model that drafts blocks of text in parallel.
Model
Rank-asymmetric flow method for pixel-space image generation, incl. AsymFLUX.2 klein finetuned from FLUX.2 klein 9B.
Model
One- and two-step image generation model distilled from Z-Image-Turbo using text-stream adversarial distillation.
Model · SupraLabs
Tiny 100M-parameter open-weights text-to-image diffusion transformer from SupraLabs, generating 256×256 images.
Model · Linum AI
2B-parameter pixel-space text-to-image diffusion model from Linum AI, trained from scratch for about $21K; Apache 2.0.
