Z-Image-Turbo: 6B Distilled S3-DiT for 8-Step Real-Time Generation
Z-Image-Turbo is Tongyi-MAI's distilled 6B image generation model achieving sub-second inference in 8 steps. Features Decoupled-DMD and DMDR distillation with RL.
Z
Z-Image-Turbo
Text-to-Image6B Parameters8-StepDistilledReal-timeZ-Image-Turbo is the distilled variant of the Z-Image family, achieving sub-second inference latency with only 8 NFEs (Number of Function Evaluations). It uses Decoupled-DMD (Distribution Matching Distillation with CFG Augmentation) and DMDR (DMD + Reinforcement Learning) for state-of-the-art few-step generation. Excels in photorealistic image generation, bilingual text rendering, and robust instruction adherence.
| Developer | Tongyi-MAI (Alibaba) |
| Release Date | 2025-12 |
| Architecture | 6B S3-DiT (Distilled via Decoupled-DMD + DMDR) |
| License | Apache-2.0 |
| Output Resolution | 512×512 to 2048×2048 (any aspect ratio) |
| Inference Steps | 8 |
| Guidance Scale | 0.0 (no CFG) |
| Text Encoder | Qwen 3 4B |
| VRAM | Fits 16G consumer GPUs |
Guides and workflows related to this model series.
No articles found.