• [2026.05.12] Technical report released.
• [2026.05.22] 1K-resolution training code, inference code, weights, and dataset released.
• [2026.05.23] Online demo. (Thanks to multimodalart for the support!)

git clone https://github.com/TencentYoutuResearch/T2I-L2P.git
cd T2I-L2P
pip install -e .

Checkpoint:

import torch
from diffsynth.pipelines.z_image_L2P import ZImagePipeline, ModelConfig

main_model_path = "/path/model-1k-merge.safetensors"

text_encoder_paths = [
    "/path/Z-Image-Turbo/text_encoder/model-00001-of-00003.safetensors",
    "/path/Z-Image-Turbo/text_encoder/model-00002-of-00003.safetensors",
    "/path/Z-Image-Turbo/text_encoder/model-00003-of-00003.safetensors",
]

tokenizer_path = "/path/Z-Image-Turbo/tokenizer"

pipe = ZImagePipeline.from_pretrained(
    torch_dtype=torch.bfloat16,
    device="cuda",
    model_configs=[
        ModelConfig(path=[main_model_path]),
        ModelConfig(path=text_encoder_paths),
    ],
    tokenizer_config=ModelConfig(path=tokenizer_path),
)

prompt = "an origami pig on fire in the middle of a dark room with a pentagram on the floor"

image = pipe(
    prompt=prompt,
    seed=42,
    rand_device="cuda",
    num_inference_steps=30,
    cfg_scale=2.0,
    height=1024,
    width=1024,
)

image.save("example.png")

First, install gradio:

pip install gradio

Launch a multi-GPU web UI:

python app.py

The demo auto-detects free GPUs, dispatches each request to an idle device, and exposes a Gradio interface at http://0.0.0.0:23231.

The full training pipeline consists of four steps: (1) prepare the Z-Image base weights → (2) convert them into a pixel-space initialization → (3) launch training → (4) merge the trained delta back with the pixel-init weights for inference.

Download the official Z-Image-Turbo checkpoint from Hugging Face:

• 🤗 Tongyi-MAI/Z-Image-Turbo

Convert the latent-space DiT weights into a pixel-space initialization that L2P can fine-tune from:

python examples/z_image/L2P_convert_weight.py \
  --latent_ckpt_files \
    /path/to/Z-Image-Turbo/transformer/diffusion_pytorch_model-00001-of-00003.safetensors \
    /path/to/Z-Image-Turbo/transformer/diffusion_pytorch_model-00002-of-00003.safetensors \
    /path/to/Z-Image-Turbo/transformer/diffusion_pytorch_model-00003-of-00003.safetensors \
  --output_path ./pretrain_weight/Z-Image-Pixel-Init/diffusion_pytorch_model.safetensors

Standard training :

bash train_run.sh

Low-VRAM training (single GPU < 24 GB VRAM):

bash train_run_low_VRAM.sh

Provide a directory of images plus a CSV metadata file:

data/
├── images/                # raw image folder
└── metadata.csv           # columns: file_name, text, ...

python merge_weights.py \
  --file_a ./models/train/L2P_Standard/step-xxx.safetensors \
  --file_b ./pretrain_weight/Z-Image-Pixel-Init/diffusion_pytorch_model.safetensors \
  --file_out ./models/train/L2P_Standard/model-merge.safetensors

• --file_a: trained checkpoint from Step 3
• --file_b: pixel-init weights from Step 2
• --file_out: merged single-file weight

If you find this work useful, please consider citing:

@article{chen2026l2p,
  title   = {L2P: Unlocking Latent Potential for Pixel Generation},
  author  = {Chen, Zhennan and Zhu, Junwei and Chen, Xu and Zhang, Jiangning and
             Chen, Jiawei and Zeng, Zhuoqi and Zhang, Wei and Wang, Chengjie and
             Yang, Jian and Tai, Ying},
  journal = {arXiv preprint arXiv:2605.12013},
  year    = {2026}
}

@article{chen2025dip,
  title   = {DiP: Taming Diffusion Models in Pixel Space},
  author  = {Chen, Zhennan and Zhu, Junwei and Chen, Xu and Zhang, Jiangning and
             Hu, Xiaobin and Zhao, Hanzhen and Wang, Chengjie and Yang, Jian and
             Tai, Ying},
  journal = {arXiv preprint arXiv:2511.18822},
  year    = {2025}
}

L2P is built upon the excellent open-source work of DiffSynth-Studio, Z-Image.