Zen Artist is the Zen LM text-to-image product built on Z-Image (Tongyi-MAI, Apache 2.0) — an undistilled single-stream diffusion transformer with a Qwen3 text encoder.

This repository is scaffolding: configuration, examples and attribution. It ships no weights. Pull the base model from the upstream repository.

Key Capabilities

Text-to-Image Generation: Generate high-fidelity images from natural language prompts
Full Classifier-Free Guidance: The base is undistilled, so CFG and negative prompting work
Finetunable: A suitable base for LoRA training and structural or semantic conditioning

Image editing is a separate product — see zen-artist-edit.

Model Specifications

Attribute	Value
Base model	`Tongyi-MAI/Z-Image` (Apache 2.0)
Transformer	`ZImageTransformer2DModel`, ~6.2B params (30 layers, dim 3840)
Text encoder	Qwen3, ~4.0B params
Pipeline	`ZImagePipeline` (diffusers)
Resolution	512 x 512 to 2048 x 2048, any aspect ratio
Guidance scale	3.0 - 5.0
Inference steps	28 - 50
License	Apache 2.0

Quick Start

Install dependencies:

pip install torch git+https://github.com/huggingface/diffusers

Download the base model:

pip install -U huggingface_hub
hf download Tongyi-MAI/Z-Image

Text-to-Image Generation

import torch
from diffusers import ZImagePipeline

pipe = ZImagePipeline.from_pretrained(
    "Tongyi-MAI/Z-Image",
    torch_dtype=torch.bfloat16,
    low_cpu_mem_usage=False,
)
pipe.to("cuda")

image = pipe(
    prompt="A cozy coffee shop at night, warm lighting, chalkboard sign with elegant lettering",
    negative_prompt="",
    height=1024,
    width=1024,
    num_inference_steps=50,
    guidance_scale=4.0,
    generator=torch.Generator("cuda").manual_seed(42),
).images[0]

image.save("output.png")

Multi-GPU Deployment

export NUM_GPUS_TO_USE=4
export TASK_QUEUE_SIZE=100
export TASK_TIMEOUT=300

Hardware Requirements

The base weights are ~12.3 GB (transformer) + ~8.0 GB (text encoder) in bf16, so a 24 GB card holds the full pipeline. Below that, use the diffusers CPU-offload and quantization paths.

Community Support

Zen Artist is compatible with the HuggingFace diffusers ecosystem, including LoRA workflows, quantization, and layer-by-layer offloading for low-VRAM deployments.

License

Apache 2.0.

The base model, Tongyi-MAI/Z-Image, is Apache 2.0. Its model card is retained verbatim at base-model/README.md. Portions of src/examples/ come from Qwen-Image, also Apache 2.0. See NOTICE.

Citation

@misc{zenlm2025zen-artist,
    title={Zen Artist: Text-to-Image Generation},
    author={Hanzo AI and Zoo Labs Foundation},
    year={2025},
    howpublished={\url{https://github.com/zenlm/zen-artist}}
}

Zen LM by Hanzo AI - Clarity Through Intelligence
zenlm.org | HuggingFace | GitHub