Z-Image Turbo

'The step-distilled 6B member of Z-Image: a single-stream DiT that makes images in about 8 steps without CFG. Training it directly breaks the distillation, so ai-toolkit trains it through a de-distilling training adapter that is merged in for training and removed for sampling.'

weights: Tongyi-MAI/Z-Image-Turbo ↗
org: Tongyi-MAI (Alibaba)
modality: image
tasks: text-to-image
license: Apache 2.0
released: 2025-11-25not verified
native output: 8 steps, no CFG
total params: 10.26B
model.arch: zimage:turbo

Components

role	model	params	size	dtype	trained
Transformer	Z-Image Turbo DiT (Diffusers) Tongyi-MAI/Z-Image-Turbo/transformer diffusers.ZImageTransformer2DModel Stored as fp32 in the Diffusers repo. ai-toolkit supplies the config from here but takes the weights from the ComfyUI file below.	6.15B 6,154,908,736	24.62 GB	fp32	yes
Transformer (loaded)	Z-Image Turbo, ComfyUI bf16 filealternate file Comfy-Org/z_image_turbo/split_files/diffusion_models/z_image_turbo_bf16.safetensors diffusers.ZImageTransformer2DModel What ai-toolkit downloads when name_or_path is Tongyi-MAI/Z-Image-Turbo and qtype is qfloat8 or quantization is off. With a convrot8 qtype it prefers z_image_turbo_int8_convrot.safetensors from the same folder. The fused ComfyUI qkv is split into to_q/to_k/to_v on load.	6.15B 6,154,908,736	12.31 GB	bf16	yes
Text encoder	Qwen3-4B Tongyi-MAI/Z-Image-Turbo/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-Turbo/tokenizer transformers.Qwen2Tokenizer	—	—	—	—
VAE	FLUX.1 VAE Tongyi-MAI/Z-Image-Turbo/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
Training adapter	Z-Image Turbo training adapter v2 (LoRA, rank 64)adapter ostris/zimage_turbo_training_adapter/zimage_turbo_training_adapter_v2.safetensors A de-distillation LoRA trained on Turbo outputs. ai-toolkit merges it into the transformer before training and subtracts it while sampling, so your LoRA learns only the new concept and still runs at 8 steps without it. v1 (half the size) is in the same repo.	170.1M 170,065,920	340 MB	bf16	no
total		10.26B	32.83 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: Step-distilled for ~8 steps and RL-tuned; runs without CFG
Objective: Rectified flow, shift 3.0
Norm / position: RMSNorm, QK RMSNorm, 3-axis RoPE [32, 48, 48], theta 256

In AI Toolkit

model.arch: zimage:turbo
UI label: Z-Image Turbo (w/ Training Adapter) (image)
model.name_or_path: Tongyi-MAI/Z-Image-Turbo
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, model.assistant_lora_path

UI defaults

quantize / quantize_te: true / true
qtype: qfloat8
low_vram: true
assistant_lora_path: ostris/zimage_turbo_training_adapter/zimage_turbo_training_adapter_v2.safetensors
train.unload_text_encoder: false
noise_scheduler / sampler: flowmatch / flowmatch
timestep_type: weighted
sample guidance_scale / steps: 1 / 9
network.conv: disabled (linear LoRA only)

Specifics

Arch name: zimage:turbo is a UI preset. The config loader strips everything after the colon, so the job runs as arch zimage with these defaults.
Training adapter: The adapter is merged into the full-precision transformer at weight 1.0 before quantization, then inverted (multiplier −1) during sampling so samples show the distilled model plus your LoRA. qfloat8 is switched to float8 when an adapter is used.
Transformer source: For the Tongyi-MAI/Z-Image-Turbo repo id, weights come from Comfy-Org/z_image_turbo (bf16, or int8_convrot for a convrot8 qtype); a local copy under MODELS_PATH wins over a download. Set model_kwargs.use_comfy_weights: false to load the Diffusers folder.
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, so 1 means no CFG. Flowmatch Euler, shift 3.0.
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_turbo_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-Turbo"        arch: "zimage:turbo"        assistant_lora_path: "ostris/zimage_turbo_training_adapter/zimage_turbo_training_adapter_v2.safetensors"        quantize: true        qtype: "qfloat8"        quantize_te: true        low_vram: true      sample:        sampler: "flowmatch"        sample_every: 250        width: 1024        height: 1024        guidance_scale: 1        sample_steps: 9        prompts:          - "a bear building a log cabin in the snow covered mountains"