datasets
Training and evaluation data, with the modality, task and licence stated up front. Listed live from the Hugging Face Hub.
sd-z-image-turboz_image_lorasZImageTurboGen
Z-Image-Turbo Synthetic Dataset (3000 Image Pairs)
Dataset Preview
Overview
This dataset contains 3,000 synthetically generated image-text pairs created using Tongyi-MAI/Z-Image-Turbo (v1.0). Each entry consists of:
A high-resolution image (zimage_{idx}.webp)
A corresponding text prompt file (zimage_{idx}.txt)
Designed specifically for advanced diffusion model research, this dataset enables:
Model de-distillation (reversing knowledge distillation effects)… See the full description on the dataset page: https://huggingface.co/datasets/lrzjason/ZImageTurboGen.Z-Image-XLA-Cache-Promptz-image-ethnicity-test
Z-Image Turbo Ethnicity Benchmarking Dataset
Overview
This dataset was created to evaluate and test the Z-Image Turbo model's capabilities in accurately rendering various ethnicities. It comprises photorealistic portrait prompts designed to cover a diverse range of ethnic groups and demographic attributes.
Generation Methodology
The prompts in this dataset were synthetically generated using the Mistral-Small-3.2-24B-Instructmodel. The… See the full description on the dataset page: https://huggingface.co/datasets/k-mktr/z-image-ethnicity-test.z-image-models# Z-Image Models for ComfyUI
Chinese style portrait generation models.
Contents
Main model: Z-Image-Base-8steps-White_Marble-AIO_v2-fp8 (9.6GB)
LoRAs: 0 files
Workflows: Base, White Jade V2, Yaoguang, Rumeng
Usage
# Download in Colab
!wget -O model.safetensors https://huggingface.co/datasets/wind1/z-image-models/resolve/main/checkpoints/Z-Image-Base-8steps-White_Marble-AIO_v2-fp8.safetensors
Uploaded: 2026-05-31 04:18:09
z-image-examples
Z-Image Turbo Portrait Dataset
This dataset contains 126 portrait prompts and their corresponding image outputs, demonstrating the capabilities of the Z-Image Turbo text-to-image model.
Model Information
Model Name: Z-Image Turbo
Hugging Face Repository: Tongyi-MAI/Z-Image-Turbo
Dataset Contents
prompts.jsonl: A JSONL file containing the 126 text prompts used for generation. Each entry includes a unique ID and the prompt text.
outputs/: Directory containing… See the full description on the dataset page: https://huggingface.co/datasets/k-mktr/z-image-examples.zimage-base-vs-turbo-comparison
⚡ Z-Image Base vs Turbo 效果对比
使用相同的 prompt 和 seed，对比 Z-Image (Base) 和 Z-Image-Turbo 的生成效果。
📊 模型信息对比
特性
Z-Image (Base)
Z-Image-Turbo
模型
Tongyi-MAI/Z-Image
Tongyi-MAI/Z-Image-Turbo
参数量
6B
6B (蒸馏版)
架构
S3-DiT (单流 DiT)
S3-DiT (单流 DiT)
推理步数
28-50 步
8 步
CFG 引导
✅ (scale=3.0-5.0)
❌ (scale=0.0)
负向提示词
✅ 支持
❌ 不支持
生成多样性
⭐⭐⭐ 高
⭐⭐ 低
视觉质量
⭐⭐⭐ 高
⭐⭐⭐⭐ 非常高
RL 强化学习
❌
✅ (DMDR)
可微调 (LoRA等)
✅
❌
VRAM 需求
~16GB+
≤16GB
🖼️ 并排对比
Prompt 1: 🐆… See the full description on the dataset page: https://huggingface.co/datasets/DealayLomoi/zimage-base-vs-turbo-comparison.ZImage-Turbo-200k-multires-aspectbucketed
ZImage-Turbo WebDataset
Generated images from the ZImage-Turbo model using DiffusionDB prompts.
Generation Details
Hardware: 8x NVIDIA RTX 3090 GPUs
Generation Time: ~2 days
Estimated Cost: ~$70 (cloud compute)
Dataset Statistics
Total Samples: 211,081
Total Shards: 216
Samples per Shard: ~1000
Shard Naming Convention
Tarballs are named: {base_resolution}-{aspect_ratio}-{shard_num:04d}-of-{total_shards:04d}.tar
For example:… See the full description on the dataset page: https://huggingface.co/datasets/RareConcepts/ZImage-Turbo-200k-multires-aspectbucketed.Z_Image_Workflowszimage-prompts-500kZ-Image-XLA-CacheHQ-Z-Image-200KZ-image-baseXiang_Z_Image_Turbo_different_style_Images
Use lora from https://huggingface.co/svjack/Z_Image_Turbo_Xiang_Lora
Xiang different style Grid Image
Z-Image-XLA-Cache-V2z-image-turbo-genXiang_Z_Image_Turbo_HairCut_ZH_Repose_Captioned
for_zimage_trainingPrince_Xiang_Z_Image_Turbo_different_style_Images
Use lora from https://huggingface.co/svjack/Prince_Xiang_Z_Image_Turbo_Lora
Xiang different style Grid Image
Prince_Xiang_Z_Image_Turbo_glasses_Images
Use lora from https://huggingface.co/svjack/Prince_Xiang_Z_Image_Turbo_Lora
Xiang Glasses Grid Image
Z-Image-XLA-Cache-V3z-image-sdxl-benchmark
Z-Image vs SDXL-Lightning Benchmark Dataset
Benchmark comparing Z-Image-Turbo (6B parameter multilingual diffusion model) against SDXL-Lightning (4-step distilled SDXL).
Key Results
Model
Latency
Memory
Languages
SDXL-Lightning
1.32s
9.6 GB
English
Z-Image-Turbo
16.51s
23.3 GB
9 languages
SDXL-Lightning is 12.5x faster, but Z-Image supports multilingual text rendering.
Dataset Contents
80 generated images (1024x1024)
50 quality benchmark… See the full description on the dataset page: https://huggingface.co/datasets/RyeCatcher/z-image-sdxl-benchmark.Z-IMAGE-TURBO-ANIMEzimage-te-cachez-image-training-datasetzimage-R
zimage-R
Prebuilt NF4 artifact of the Z-Image-Turbo transformer for the
diffuseR R package.
Built with diffuseR::flux_quantize(format = "nf4") from
Tongyi-MAI/Z-Image-Turbo
(Apache-2.0): NF4-packed uint8 weights with float32 absmax blocks,
bfloat16 residents, sharded under 2 GB so the CRAN release of the
safetensors R package reads it.
Fetched automatically by diffuseR::download_zimage_turbo() when the
resolved precision is nf4. The text encoder, VAE, and tokenizer come
from the… See the full description on the dataset page: https://huggingface.co/datasets/cornball-ai/zimage-R.Z-Image-Turbo-Male-Model-ImagesQFabric_mt_images_1024Xiang_POV_Touch_Z_Image_Turbo_Images
Use lora from https://huggingface.co/svjack/Z_Image_Turbo_Xiang_Lora
Xiang POV Touch Grid Image