Z-Image

'The undistilled 6B foundation model of the Z-Image family: a single-stream DiT on Qwen3-4B text features and the FLUX.1 VAE, with full CFG and negative prompts. It is the Z-Image checkpoint meant for fine-tuning and trains directly, with no adapter.'

weights: Tongyi-MAI/Z-Image ↗
org: Tongyi-MAI (Alibaba)
modality: image
tasks: text-to-image
license: Apache 2.0
released: 2026-01-23not verified
native output: 512×512 to 2048×2048 (any aspect), 28–50 steps, CFG 3–5
total params: 10.26B
model.arch: zimage

Components

role	model	params	size	dtype	trained
Transformer	Z-Image DiT Tongyi-MAI/Z-Image/transformer diffusers.ZImageTransformer2DModel Same shape as Z-Image Turbo. ai-toolkit loads this Diffusers folder directly (no ComfyUI file is registered for this repo).	6.15B 6,154,908,736	12.31 GB	bf16	yes
Text encoder	Qwen3-4B Tongyi-MAI/Z-Image/text_encoder transformers.Qwen3ForCausalLM 36 layers, 2560 hidden, tied embeddings (151,936 vocab). Only the second-to-last hidden state is used.	4.02B 4,022,468,096	8.04 GB	bf16	no
Tokenizer	Qwen3 BPE tokenizer Tongyi-MAI/Z-Image/tokenizer transformers.Qwen2Tokenizer	—	—	—	—
VAE	FLUX.1 VAE Tongyi-MAI/Z-Image/vae diffusers.AutoencoderKL Same config as the FLUX.1 VAE (scaling 0.3611, shift 0.1159).	83.8M 83,819,683	168 MB	bf16	no
total		10.26B	20.52 GB

Latent space

spatial

8×

channels

patch

2×2

autoencoder: FLUX.1 VAE
pixels per token: 16×16
notes: The FLUX.1 latent space. The transformer patchifies 2×2 internally (all_patch_size [2]), so one token covers 16×16 pixels.

input	latent (c×t×h×w)	tokens
512×512	16×1×64×64	1,024
1024×1024	16×1×128×128	4,096	ai-toolkit sample default
2048×2048	16×1×256×256	16,384

Architecture

Blocks: 30 single-stream blocks, plus 2 noise-refiner blocks (image only) and 2 context-refiner blocks (text only)
Hidden size: 3840 (30 heads × 128)
FFN size: 10240 (SwiGLU)
Text conditioning: Qwen3-4B second-to-last hidden states (2560-d) through the chat template, up to 512 tokens, padding dropped. Text and image tokens share one sequence
Timestep: adaLN modulation from a 256-d timestep embedding (context refiner is unmodulated)
Distillation: None. Full CFG and negative prompts
Objective: Rectified flow, shift 6.0 in the repo scheduler config
Norm / position: RMSNorm, QK RMSNorm, 3-axis RoPE [32, 48, 48], theta 256

In AI Toolkit

model.arch: zimage
UI label: Z-Image (image)
model.name_or_path: Tongyi-MAI/Z-Image
source: extensions_built_in/diffusion_models/z_image/z_image.py
toolkit/models/v2/diffusion_models/z_image.py
toolkit/models/v2/text_encoders/qwen3.py
toolkit/models/v2/vae/autoencoder_kl.py
toolkit/models/v2/resolver.py
toolkit/config_modules.py
toolkit/models/registry.py
extensions_built_in/diffusion_models/ui.tsx
extra UI sections: model.low_vram, model.layer_offloading

UI defaults

quantize / quantize_te: true / true
qtype: qfloat8
low_vram: true
train.unload_text_encoder: false
noise_scheduler / sampler: flowmatch / flowmatch
timestep_type: weighted
sample_steps: 30
network.conv: disabled (linear LoRA only)

Specifics

Transformer source: transformer/ from name_or_path. A single .safetensors file also works (config from Tongyi-MAI/Z-Image-Turbo, which has the same shape).
Training adapter: Not needed: the model is undistilled. assistant_lora_path is still accepted and handled like Turbo’s.
Text encoder and VAE source: text_encoder/, tokenizer/ and vae/ from extras_name_or_path (defaults to name_or_path). Single .safetensors checkpoints fall back to Tongyi-MAI/Z-Image-Turbo.
Resolution: Buckets snap to multiples of 16 (8× VAE × 2×2 patch).
Timesteps: The model takes t = (1000 − timestep) / 1000 (1 = clean) and its output is negated to match the noise − latents flow target.
Quantization: t_embedder, cap_embedder, all_x_embedder and all_final_layer stay in full precision.
Sampling: guidance_scale is shifted down by 1 before it reaches the pipeline (the UI’s 4 becomes 3). Flowmatch Euler with ai-toolkit’s fixed shift 3.0, not the repo’s 6.0; training uses the same 3.0 scheduler.
LoRA target: ZImageTransformer2DModel
Saving: Full fine-tunes save one ComfyUI-layout file (fused qkv). LoRA keys use the ComfyUI diffusion_model prefix.
Metadata base version: zimage

Example config

not verified

job: extensionconfig:  name: "my_zimage_lora_v1"  process:    - type: "sd_trainer"      training_folder: "output"      device: cuda:0      network:        type: "lora"        linear: 32        linear_alpha: 32      save:        dtype: bf16        save_every: 250        max_step_saves_to_keep: 4      datasets:        - folder_path: "/path/to/images"          caption_ext: "txt"          caption_dropout_rate: 0.05          cache_latents_to_disk: true          resolution: [512, 768, 1024]      train:        batch_size: 1        steps: 3000        gradient_checkpointing: true        noise_scheduler: "flowmatch"        timestep_type: "weighted"        optimizer: "adamw8bit"        lr: 1e-4        dtype: bf16      model:        name_or_path: "Tongyi-MAI/Z-Image"        arch: "zimage"        quantize: true        qtype: "qfloat8"        quantize_te: true        low_vram: true      sample:        sampler: "flowmatch"        sample_every: 250        width: 1024        height: 1024        guidance_scale: 4        sample_steps: 30        prompts:          - "a bear building a log cabin in the snow covered mountains"