Tencent YouTu Research
Tencent's pixel-space image generator — no VAE artifacts, native 1K quality
L2P (Latent-to-Pixel) from Tencent YouTu Research bypasses the VAE compression bottleneck of standard diffusion models, generating images directly in pixel space. By freezing a 6B-parameter latent diffusion backbone and training only shallow adapter layers, it transfers all learned priors into native 1024×1024 pixel generation with no latent-space color shift or blur. Weights on HuggingFace; ComfyUI nodes available for drop-in use. **Step-by-step:** 1) `git clone https://github.com/TencentYoutuResearch/T2I-L2P`. 2) Download weights: `huggingface.co/zhen-nan/L2P`. 3) `pip install -r requirements.txt`. 4) Run the inference script with a text prompt and target resolution. 5) Alternatively, install the ComfyUI nodes (rebels-l2p or gjnave/ggf-ltp-zimage). **VRAM:** 8 GB (512px) · 24 GB recommended for native 1K.
Pixel-space generation eliminates the VAE color-shift and blurriness of latent models.
Free (open-source, self-hosted) · ~20 GB model on disk