Z-Image Turbo
Model background
- Developer: Alibaba Tongyi-MAI lab. License: Apache 2.0 (commercial use permitted).
- Architecture: 6.15B parameter Single-Stream Diffusion Transformer (S3-DiT) — not based on Flux or SDXL.
- Text encoder: Qwen3-4B base (not Qwen3-VL or Qwen3-Thinking — using others is counterproductive).
- Speed: ~0.8s for 8-step 512×512 on H800; distilled via Decoupled-DMD + DMDR.
- Benchmark: Ranked #1 open-source at launch (Dec 2025); Flux.2 variants have since overtaken it on open weights. (source: z-image-turbo.md)
Variants
| Variant | Steps | CFG | Status |
|---|---|---|---|
| Z-Image-Turbo | 8 | 1.0 | Released Nov 26, 2025 |
| Z-Image (Base) | 28–50 | 3.0–5.0 | Released Jan 27, 2026; supports negative prompts |
| Z-Image-Omni-Base | 50 | on | Upcoming |
| Z-Image-Edit | 50 | on | Upcoming |
Optimal settings
Draw Things CLI (authoritative)
| Parameter | Value | config-json key |
|---|---|---|
| Steps | 8 | `"steps": 8` |
| CFG | 1.0 | `"guidanceScale": 1` |
| Sampler | DDIM Trailing | `"sampler": 15` |
| Resolution | 1024×1024 (~2MP sweet spot; up to 4MP) | `"width"` / `"height"` |
| Aspect ratios | 1:1, 3:4, 9:16, 16:9 | |
| Negative prompt | Not supported — ignored |
See Draw Things CLI — Sampler Integer IDs for the full sampler integer table. Note: the community catalog default is `"sampler": 17` (UniPC AYS); DDIM Trailing (`15`) is this project's preference.
Reference (Diffusers / ComfyUI)
| Parameter | Value | Notes |
|---|---|---|
| Steps | 9 | = 8 effective DiT forwards |
| guidance_scale | 0.0 (Diffusers) / 1.0 (ComfyUI KSampler) | Do not raise |
| Sampler | euler / res_multistep (official) |
Prompt anatomy
Z-Image Turbo reads prompts as natural-language sentences. Midjourney/SD-style keyword lists underperform. Write 80–250 words using this 6-part structure, with subject and any text to render placed first: (source: z-image-turbo.md)
1. Subject — who/what, age, clothing, materials, at least one non-idealized feature for people 2. Scene — location, time of day, weather, 1–2 supporting props max 3. Composition / camera — shot type (close-up, medium, wide), angle, lens/focal length 4. Lighting — direction, color temperature, quality (single highest-impact variable) 5. Style — ONE dominant medium (photoreal / cinematic / oil painting / cel-shaded anime); never mix conflicting styles 6. Positive constraints — what must be present, phrased as presence not absence
Rules of thumb:
- 80–250 words is the sweet spot; beyond ~300 words coherence degrades
- Limit to 3–5 strong visual concepts — overloading causes contradictions
- Write like a creative brief or shot list, not poetry
Negative prompts don't work
Z-Image Turbo runs at guidance_scale 0, so negative prompts are completely ignored. Convert every "don't" into a positive constraint: (source: z-image-turbo.md)
| Instead of (negative) | Use (positive) |
|---|---|
| blurry | sharp focus, crisp details, clean detailed image |
| extra fingers | correct human anatomy, natural hands and fingers |
| cluttered | simple uncluttered background, clean seamless backdrop |
| text/watermark | no text, no watermark, no logos (modern text encoders still parse "no X" semantics) |
Z-Image Base (the non-distilled variant) does support negative prompts at CFG 3–5.
Fixing the plastic face problem
The default prior is glossy and airbrushed. "Realistic" barely helps. The fix is photography vocabulary: (source: z-image-turbo.md)
- Pick ONE camera phrase: `"shot on a point-and-shoot film camera"`, `"Canon R5 with 85mm f/1.4"`, `"Leica M6"`, `"Fujifilm GFX100 II medium format"`
- Pick ONE film stock: `"Kodak Portra 400 tones"`, `"Cinestill 800T halation"`, `"Ilford HP5 black-and-white grain"`
- Add texture words: `"visible pores"`, `"fine skin texture"`, `"slightly asymmetrical features"`, `"film grain"`
- Add realism cues: `"candid, unstaged, documentary"`, `"realistic, ordinary, everyday appearance, not a model"`
Stacking more than one camera phrase or film stock dilutes the effect.
Text rendering
Z-Image Turbo excels at rendering legible text (including CJK): (source: z-image-turbo.md)
- Put exact strings in quotes: `a sign that reads "GARDEN OF LIFE"`
- Describe font style: `"bold sans-serif"`, `"calligraphic brush"`
- Describe position: `"centered, top"`
- Add: `"legible, sharp edges"`
- For bilingual: include both quoted strings — keep CJK strings short and common for best accuracy
Quality-boosting tokens
Unlike SD-era "masterpiece, 8k, trending on artstation," Z-Image responds to concrete photographic specifics: (source: z-image-turbo.md)
- Camera/lens: `"shot on Leica M6"`, `"Canon R5 with 85mm f/1.4"`, `"Sony A7R V, 24mm"`
- Film stock: `"Kodak Portra 400 tones"`, `"Cinestill 800T halation"`, `"Fujifilm Velvia 50"`
- Texture: `"visible pores"`, `"fabric detail"`, `"imperfections"`, `"film grain"`
- Lighting: `"soft diffused overcast light"`, `"three-point studio lighting"`, `"golden hour"`, `"rim light"`
- Composition: `"rule of thirds"`, `"shallow depth of field"`, `"ultra-wide"`
Things to avoid
- Mixing conflicting styles ("photorealistic cartoon")
- Vague terms ("beautiful, nice")
- SD-era quality spam ("masterpiece, 8k, trending on artstation")
- Relying on the negative-prompt box
- Over-long poetic prompts
- Stacking multiple cameras or film stocks in one prompt
(source: z-image-turbo.md)