~8-step distill — iterate prompts aggressively
Decoupled-DMD / DMDR few-step distillation is the official pitch: composition, expression, wardrobe can be tested fast; waiting half a minute per tweak kills short-video cover schedules.
Alibaba Tongyi MAI · Z-Image-Turbo · 2025-11
Z Image Turbo — Tongyi’s 6B open realism workhorse
~8-step distilled inference, photoreal skin and light, readable Chinese/English in-image text — use free credits on iMini for avatars, headshots, and bilingual poster drafts first.
- Free credits
- ~8-step Turbo
- Photoreal realism
- Bilingual in-image text
What is Z Image Turbo?
Z Image Turbo is Tongyi MAI (Tongyi Lab)’s distilled text-to-image model open-sourced in November 2025 — a ~6B single-stream diffusion Transformer (S3-DiT) that ships at ~8 NFE steps on paper, sub-second latency on enterprise H800, and runs on ~16GB consumer GPUs. Official benchmarks put it against closed flagships on photoreal detail, light, materials, and bilingual in-image text (small poster type still readable while faces hold). GitHub / Hugging Face also cite #1 open model on Artificial Analysis text-to-image and strong Elo on Alibaba AI Arena vs closed models. Family split: Turbo for fast generation, Edit for instruction edits. On iMini it is the default for free-credit realism / portrait / bilingual drafts — switch to Qwen Image Plus for long Chinese poster layout, Nano Banana Pro / GPT Image 2 for native 4K polish. Blurry skin, fake light, or broken shop-sign Chinese kills headshots and bilingual KV — Turbo’s job is to cheap-test those failures before final art.
Developer
Alibaba Tongyi MAI
Open-sourced
November 2025
Scale / architecture
~6B · S3-DiT
Inference
~8 steps · sub-second on H800
Strengths
Photoreal realism · CN/EN text
On iMini
Free credits · browser
Photography-grade skin and light
Official chapter on detail, light, texture, mood — avatars, headshots, makeup close-ups pass the “looks human” gate first; one fake face voids the whole portrait set.
Bilingual in-image text while keeping faces
Official emphasis on bilingual rendering rivaling top closed models; small poster type still holds. Wrong shop copy or blurred titles kill bilingual KV / campaign visuals.
Start on iMini free credits
Open weights (Apache 2.0), enter workspace from this page with free credits — no vendor key setup first.
What the model actually delivers
Car-window rim light, indoor window light, gown side light, soft beauty light, Peking opera makeup, creative mask — official photoreal samples where skin and light hold up.
How it compares to mainstream image models
Z Image Turbo vs poster-oriented Qwen Image Plus, Google’s Nano Banana 2, and OpenAI’s GPT Image 2 — step count, photoreal portraits, bilingual text, and default picks side by side.
| Dimension | Z Image Turbo | Qwen Image Plus | Nano Banana 2 | GPT Image 2 |
|---|---|---|---|---|
| Speed profile | ~8-step distill · sub-second (H800) | Instant fast draft + quality tiers | ~8-step distill · sub-second (H800) | Instant fast draft + quality tiers |
| Portraits | Photoreal realism (official) | Good | Photoreal realism (official) | Photoreal realism (official) |
| In-image text | Good | Industry-best text precision | Photoreal realism (official) | Photoreal realism (official) |
| Resolution | Up to 2K | Up to 2K | Up to 4K | Native 4K (3840×2160) |
| Chinese support | Official strength in CN/EN in-image text | Official strength in CN/EN in-image text | Multi-reference support | Multi-reference support |
| Draft speed | ~8-step distill · sub-second (H800) | Instant fast draft + quality tiers | ~8-step distill · sub-second (H800) | Instant fast draft + quality tiers |
| Editing | Basic repaint / variations | Basic repaint / variations | Basic repaint / variations | High-fidelity edit + identity hold |
| Default pick | Default for photoreal portraits / bilingual drafts | Chinese posters and long copy | Speed + 4K daily workhorse | OpenAI default for new projects |
Production problems it actually solves
Few-step realism, portrait texture, bilingual text — cheap iteration on this lane before final art.
Few-step inference you can iterate
~8 steps per image — swap prompts, light, makeup without queue pain; slow iteration breaks sample schedules.
Photoreal portraits that pass review
Skin, hair, rim light, depth of field — avatars and headshots pass “looks human” first; one fake face voids the set.
Chinese and English text in-frame
Shop signs, poster lines, bilingual titles readable; blurred or wrong text kills campaign assets.
Materials and mood
Fabric, beads, dappled daylight — same portrait spec as official samples.
Cheap layer before final art
Burn free credits while direction is unsettled; upgrade to Nano Banana Pro or GPT Image 2 for 4K / multi-ref polish.
Realism workflows that default to Turbo first
Official samples map to these deliverables: lifestyle candid, daily headshot, editorial, beauty, makeup, creative portrait.
Lifestyle portraits
Car-window rim light, travel candids — highlights and ambient light must land; fake light fails lifestyle review.
Daily headshots
Indoor window-light half-body for social and profile drafts; fake skin or muddy hair kills profile output.
Editorial portraits
Gown, side light, depth of field — magazine half-body; collapsed light or fake face voids the editorial set.
Soft beauty portraits
Soft beauty light, lace and hair detail; muddy materials kill beauty / lookbook samples.
Makeup close-ups
Peking opera makeup and beaded headpieces — features and materials must hold; muddy detail makes close-ups useless.
Creative portraits
Masks, embroidery, unconventional styling still human; one drifted look kills creative pitches.
Who should default to it
Teams that need cheap, fast photoreal portraits and bilingual drafts that pass first review.
Creators
Covers, thumbnails, portrait samples — iterate hard on free credits first.
Marketing teams
Campaign KV and bilingual short-copy poster drafts, compared with flagships in-browser.
E-commerce
Model shots and headshot direction unsettled — lock light and expression first.
Designers
Pitch boards and look direction without the most expensive lane before sign-off.
Founders
Early brand visuals and persona art — cheap trials until the story is clear.
Agencies
Many client sample versions, compared with Qwen / Banana / GPT Image on one screen.
Three steps to start
Step 1
Open Z Image Turbo
In the iMini image workspace, select Z Image Turbo.
Step 2
Write a clear prompt
Subject, style, aspect ratio; quote in-image text.
Step 3
Generate and iterate
Download what works, or compare with sibling models on iMini.
Copy-ready prompts
Starters aligned to official photoreal samples — copy, swap subject, generate free on iMini.
Car-window lifestyle portrait
Young East Asian woman in passenger seat smiling at camera, highway guardrail and blue sky outside, strong rim light, hair highlights, real skin texture, 85mm half-body, vertical frame.
Indoor daily portrait
Long black hair young woman half-body, white printed tee, bookshelf and window light behind, natural daylight from right, real skin and hair detail, shallow depth of field, vertical frame.
Evening-gown editorial
Middle-aged East Asian man three-quarter profile, black tux and bow tie, golden hour light, lake and distant city skyline behind, cinematic shallow depth of field, magazine editorial portrait.
Soft beauty portrait
Young woman soft-light half-body, light pink lace top, long curls over shoulders, clear hair and skin detail, dappled natural light with slight prism flare, shallow depth of field, vertical frame.
Peking opera makeup close-up
Peking opera dan role half-body close-up, traditional white base red cheek makeup, ornate blue-gold beaded headdress, red embroidered costume, warm side light, blurred background, vertical photoreal portrait.
Creative mask portrait
Young East Asian woman half-body, lower face in embroidered cat mask with gold bead edge and fine stitching, long bangs, dappled tree shade daylight, neutral wall background, vertical photoreal portrait.
Align on these before you pick
Speed claims, realism, how it splits work with Qwen / flagships, and how credits work.
Long Chinese posters and dense layout lean Qwen Image Plus; photoreal portraits / realism with bilingual short copy and fast iteration default to Z Image Turbo.
Run a Tongyi photoreal portrait free
6B · ~8 steps · photoreal realism and CN/EN text — try on iMini before polish.
Generate free on iMiniFree credits to start · no separate API key