🚀

Image Generation - Z-Image-Turbo

🛠

Text-to-Photorealistic Image

Generate high-quality photorealistic images from text prompts. Z-Image-Turbo is a distilled, efficient diffusion model which supports bilingual prompts (English & Chinese), excellent text rendering, and strong instruction adherence

What it does

Z-Image-Turbo converts text descriptions into photorealistic images with exceptional speed. Using a Scalable Single-Stream Diffusion Transformer (S3-DiT) architecture, it generates high-quality images in just 8 steps with sub-second inference latency. Perfect for applications requiring fast, high-quality image generation.

Problem it solves

  • Photorealistic quality – High-quality images matching commercial competitors
  • Bilingual text rendering – Accurate English and Chinese text in images
  • Strong instruction following – Interprets prompts accurately and consistently
  • Text understanding – Can embed readable text and complex descriptions in images
  • Creative workflows – Fast iteration for design and creative applications

Input/Output

  • Input: Text prompt (English or Chinese)
    • Format: Natural language description
    • Length: Short to detailed prompts (works best with descriptive text)
    • Language: English, Chinese, or bilingual
    • Examples:
      • "A photorealistic portrait of a woman with blue eyes"
      • "一个穿着红色连衣裙的女人站在樱花树下" (A woman in a red dress standing under cherry blossoms)
      • "Modern minimalist kitchen with marble countertops and stainless steel appliances"
  • Options:
    • diffusion steps (5 ~ 30)
      • Diffusion steps control how many refinement iterations the model uses to generate an image. The model starts with pure noise and gradually removes noise at each step until reaching the final image. More steps = better quality but slower generation (default: 9)
    • random seed
    • output image height & width
  • Output: Generated image
    • Quality: Photorealistic, high detail
      • image

Accuracy & Speed

  • Quality Metrics:
    • Elo Rating: Competitive with/exceeds leading open-source models
    • Leaderboard: Top performance on Alibaba AI Arena evaluation
    • Text Rendering: 90%+ accuracy for English and Chinese text in images
    • Instruction Following: 85-90% accuracy on detailed, complex prompts
    • Photorealism Score: Exceeds competitors in realistic detail and coherence

Technical Details

Architecture
• Type: Scalable Single-Stream Diffusion Transformer (S3-DiT) • Approach: Unified input stream (text tokens + visual semantic tokens + image VAE tokens concatenated) • Efficiency: Single-stream design maximizes parameter efficiency vs. dual-stream models • Backbone: Transformer-based diffusion with optimized attention
Key Innovation - Decoupled-DMD
• Acceleration Technique: Reduces diffusion steps without quality loss • Mechanism: Decouples time-dependent features for faster convergence • Result: 8-step generation matching 20+ step models in quality
Training Data
• Trained on large-scale image-text pairs • Multilingual: English and Chinese text datasets • Diverse content: Photorealistic, artistic, text-rendering examples
Inference Settings
• Guidance Scale: 7.5 (default, adjustable 1-15) • Height/Width: 1024×1024 (supports other resolutions) • Seed: Reproducible with fixed seed • Precision: bfloat16 recommended for optimal speed

Compliance & Provenance

Provider
Open-source
Provider type
Specialized
License
Apache 2.0
EU AI Act risk class
Limited Risk
Art. 50 transparency
Required — outputs are marked. See AI Policy §2.
Region availability
Available globally
Training data summary
Pending — provider has not yet published per Art. 53(d)

For more on how we classify models and mark outputs, see our AI Policy.

Model Source