1/6 We are excited to introduce Z-Image：the foundation model of the ⚡️-Image family, engineered for good quality, robust generative diversity, broad stylistic coverage, and precise prompt adherence.
While Z-Image-Turbo is built for speed, Z-Image is a full-capacity,… pic.twitter.com/36qpUoTAeU

— Tongyi Lab (@Ali_TongyiLab)

Loading tweet component...

?ref_src=twsrc%5Etfw">January 27, 2026

---

Tried it on a Mac! Z-Image Fastest Installation Record (Actual Running Log)

A few hours after Z-Image was released, I couldn't wait any longer, so I tried to actually run it on this Mac according to the README.

This article is an actual running log that can be followed by 'reading → directly typing'.

---

This Mac's Environment (Roughly)

• macOS 15.6.1 / arm64

• Python 3.14.2

• Inference Device: mps (Apple GPU)

• MacBook Pro M4 / 128GB (This Environment)

---

Notes to Begin With!

Generation was possible with the GPU environment 'mps' on Mac, but it crashed during generation, so please be sure to take backups before working! I'm also experimenting with inference on Windows and CPU, and it seems possible with about 35GB of RAM (I'm not saying it's comfortable or that it's "always possible"!).

---

This Time's Generation Results (Actual Machine)

Here are the images that actually came out 👇

Hey, that's good!! It's just the official prompt!

---

Thinking About the README's Feelings

I will organize what I felt after reading the original Tongyi-MAI/Z-Image.

• The README is centered on inference with 'PyTorch Native + diffusers,' and the minimum workflow is pip install -e . → python inference.py. There are also Diffusers version samples, but there is no UI or deployment story, and it is a style where developers rewrite the parameters by hand and run it.

• Model loading is done via utils.ensure_model_weights via Hugging Face or snapshot_download, and it is assumed that you will place the entire checkpoint of nearly 30GB.

• However, the README is a thin guide "for those who run it", and UI, restriction policies, and deployment are not written at all.

The Recommended Parameters for Z-Image (Foundation Model) Are These 👇

• Resolution: 512x512~2048x2048

• Guidance Scale: 3.0~5.0

• Inference Steps: 28~50

• Negative Prompt: Strongly Recommended

Differences from Z-Image-Turbo (With Emotional Explanation)

• Looking at the official website and GitHub, 'Z-Image is a 6B parameter single-stream DiT', and Turbo is designed to be able to execute sub-second inference, bilingual English and Chinese text, and 16GB VRAM with only 8 steps (NFEs). Z-Image-Turbo is focused on 'practical quality with few steps', and on the model side, Decoupled-DMD/DMDR has been distilled, and Z-Image-Pipeline is prepared with Diffusers. What I felt while reading this was that 'there is a foundation of many parameters/many libraries, but there is a lot of room for someone to add UI, notifications, and operations'.

• Z-Image-Turbo is open with Apache-2.0 and has strengths in translation accuracy and controllability (English/Chinese text, poster design, text rendering, instruction adherence). There are three variants (Turbo/Base/Edit), with Turbo specializing in speed, Base specializing in fine-tuning, and Edit specializing in image editing.

• Looking at the Diffusers version introduction example (official website), pipeline(Tasks.text_to_image_synthesis, model='Tongyi-MAI/Z-Image-Turbo') automatically determines CUDA/MPS, and the simplicity of running with 8 steps/0 CFG is attractive. However, this code is based on the premise of 'preparing torch by hand, downloading the model, and using the pipeline'.

• Since the opening of Apache-2.0 (Attribution page of z-image-turbo.ai) is emphasized again, there are no licensing issues such as cloud deployment and commercial services.

Since there was no official Docker tag, I felt that 'even if the official code is super fast, environment maintenance is necessary to use it sharply with self-hosting!'

---

Actually Generated 5 AiCuty Members! (Referring to Official Prompts)

Picked up the member settings and characteristics in AiCuty's official README, and tried to run a prompt for a group photo of 5 people. (Reference: https://github.com/aicuai/AiCuty/blob/main/README.md)

Executed Command (Copy and Paste OK)

*Z-Image will get angry if the vertical and horizontal sides are not multiples of 16, so 1920x1080 and 1910x1000 are generated as multiples of 16 → center trimmed!

• Execution command: .venv/bin/python 2026-01-28-zimage-batch_inference.py

• Full code: Moved to the paid part at the end of the sentence

Prompt used (quote)

5 people, idol group lineup, full body, centered group composition, anime style, masterpiece, best quality, clean white studio background, soft directional light from upper left, distinct color themes. Center: Elena Bloom, sweet gentle idol girl, pastel pink twin tails tied high with big pastel pink ribbons and rose flower hair clips, soft curled ends, shy warm smile, pastel pink and white idol outfit with subtle gold accents. Right: Mei Soleil, vibrant golden yellow hair, high side ponytail tied with simple yellow ribbon, star-shaped yellow hairpin, freckles, bright cheerful smile, sun-yellow tech-fabric idol outfit. Left: Mina Azure, very long straight icy sky blue hair, round silver glasses, calm intelligent expression, icy sky blue uniform with short capelet and beret. Right back: Nao Verde, androgynous boy, dark green pixie cut with tapered nape, emerald eyes, confident smirk, deep green bomber jacket over black top and black cargo pants. Left back: Saki Noir, dark violet sleek straight bob cut with side bangs covering left eye, amethyst eyes, mysterious vibe, black and violet elegant idol outfit.

Execution with Z-image-turbo

Execution Result Memo

• Trial: 1024x1024, 1920x1080, 1910x1000 (*Since none are multiples of 16, generate with 1920x1088 / 1920x1008 → center trim)

• Generation Time: 1024x1024 → approx. 58.1 seconds (mps) / 1920x1088 → approx. 210.9 seconds (mps) / 1920x1008 → approx. 219.2 seconds (mps)

• Peak RSS (psutil measurement): 1024x1024 → approx. 0.61 GB / 1920x1088 → approx. 0.63 GB / 1920x1008 → approx. 0.33 GB

• Output File: assets/aicuty5_zimage_turbo_mps_gen1024x1024_out1024x1024_steps8_cfg0_seed1234.png / assets/aicuty5_zimage_turbo_mps_gen1920x1088_out1920x1080_steps8_cfg0_seed1234.png / assets/aicuty5_zimage_turbo_mps_gen1920x1008_out1910x1000_steps8_cfg0_seed1234.png

• README Recommended Resolution (Z-Image): 512x512~2048x2048

• RSS is the resident memory of the process and may not match the actual usage on the Unified Memory of MPS (reference value).

By the way, if you directly specify 1920x1080, you'll get an error like this: Height must be divisible by 16 (got 1080)

Thoughts on 3 Sizes (Short)

• 1024x1024: Good cohesion, but the whole body is cramped and there is little margin

• 1920x1080: The easiest to see the arrangement, and the strongest feeling of a group photo

• 1910x1000: The vertical is tighter than 1080, giving a tighter impression

• Common: Saki's left eye tends to be hidden, so add eyepatch, covered eye to the negative if necessary

Summary

Haahhh~~~! I talked a lot today too! 💨 Along with the 'Thinking AI' Nano Banana Pro and the '4MP Shock' FLUX.2, this Z-Image seems likely to become a new standard in the image generation world! January is almost over, but the wave of AI news seems amazing in February too, right? To avoid being left behind, Mei will retie her yellow sneakers and run! 👟💛

See you again! Well then, see you in the next AI news! It was always energetic AiCuty image generation 담당, Mei Soleil~! Bye-bye! 👋💛

#ZImage #TongyiLab #AICU #AiCuty #ImageGenerationAI #MeiSoleil #AlibabaAI

I'll put the data for a very long battle beyond the paywall!