Uncensored Z-Image Turbo: Setup, Low-VRAM Files and Prompt Playbook

Z-Image Turbo in thirty seconds

Maker

Alibaba Tongyi Lab (Tongyi MAI)

Public since

November 2025

Size

About 6.15B, S3-DiT

Sampling

8 steps, guidance off

Lettering

Chinese and English

Open weights?

Yes, Apache-2.0

Hardware

Consumer GPUs, under 16 GB

Uncensored edition?

None official

Where Z-Image Turbo comes from

This is the latest speed-tuned model in the Z-Image series from Alibaba's Tongyi Lab, built by the Tongyi MAI team. It was also the first of the series to go public: the Hugging Face repository is dated November 25, 2025 and the GitHub news post November 26. Under the hood sits a Scalable Single-Stream Diffusion Transformer (S3-DiT) of roughly 6.15 billion parameters, where prompt tokens, visual semantic tokens and image latents travel through one shared sequence. Distillation cut sampling down to 8 function evaluations with guidance switched off; the team quotes sub-second renders on an H800 data-center GPU and says a 16 GB consumer card is enough. Alibaba makes the weights available on Hugging Face and ModelScope, and a hosted endpoint on Alibaba Cloud Model Studio is simply one more way in.

Can you get an uncensored Z-Image Turbo?

Two things drive the question. First, the hosted endpoint on Alibaba Cloud screens every prompt and every finished image, so boudoir shots, pin-ups and other adult art get rejected there. Second, the Apache-2.0 weights run on your own machine with no server in the loop, and Hugging Face now lists many community LoRAs trained on top of them. So the honest answer is mixed: there is no official uncensored Z-Image Turbo, and what a local copy will draw depends on its training data rather than on a hidden switch. Unofficial NSFW fine-tunes exist, some built on real celebrities, and we do not link them, because the license covers code and weights, not anyone's face. If you would rather skip a 20 GB download, Spicy AI's generator runs Seedream, Krea 2 and Wan 2.6 without content filters for adults (18+); it is a stand-in, not Turbo itself.

First image in fifteen minutes: the ComfyUI route

Most people never touch the official repository. The Comfy-Org repack had roughly 9.3 million downloads over 30 days in October 2026, against about 654,000 for the Tongyi-MAI original, because the original stores the transformer in F32 at about 24.6 GB. A working install always has three pieces, and missing the text encoder is the classic first-run error. Grab these, drop them into the folders below and open the official ComfyUI template.

• Loaders: UNETLoader for the model, CLIPLoader set to type lumina2 for the encoder, VAELoader for the VAE.
• Template values worth keeping: 1024×1024, 8 steps, CFG 1.0, res_multistep sampler, simple scheduler, ModelSamplingAuraFlow shift 3.
• The negative branch is zeroed by ConditioningZeroOut, so anything typed there is ignored.
• Raising CFG to 7 or steps to 30 does not help Turbo; it burns colors or just costs time.
• For pose, depth or edge control, Alibaba PAI publishes Z-Image-Turbo-Fun-Controlnet-Union-2.1 as a separate download.

"Qwen3-4B class" is inferred from the ComfyUI file name and its config; the model card does not name the encoder in one line.

Short on VRAM? Match the file to your card

Memory figures for this model come from three separate places, and each assumes different software, so read them side by side rather than as one promise. The Tongyi team says the model runs comfortably below 16 GB. DiffSynth-Studio documents an 8 GB setup with CPU offload and warns that FP8 quantization visibly softens results. A stable-diffusion.cpp wiki page describes a community recipe for 4 GB cards using Q3 or Q4 GGUF files plus offloading, at a clear cost in speed. Remember that a GGUF file replaces only the diffusion model: the text encoder and VAE still have to be downloaded.

Cache-DiT, LeMiCa and MeanCache advertise extra speed-ups; those numbers come from the community projects, not from Alibaba.

Eight steps or fifty? Turbo next to its slower sibling

The base model that arrived in January 2026, nicknamed Base, shares Turbo's architecture and text encoder but keeps classifier-free guidance and negative prompts. That makes the choice less about quality and more about workflow. If you iterate fast and want a polished photo look, stay on Z-Image Turbo. If you need a cast of distinct faces, illustration styles or a model to fine-tune, switch to Base and accept roughly four to six times the sampling time.

Variety and fine-tuning labels are the team's own ratings in the official model table, not an independent test.

Prompting a model that ignores negative prompts

With guidance off, the model only listens to what you ask for, so the positive prompt carries the whole job. Write it the way a photographer briefs a shoot, in full sentences and in a steady order: who, how they look, what they do, where, under which light, through which lens, in what style. Long prompts are fine; the hosted API caps them at 800 characters, a sensible ceiling locally too. The comparison below shows how much a vague prompt leaves to chance.

Five Z-Image Turbo prompts to start from

Fill in the brackets and keep the order of the parts. All five assume adult subjects. They work in the ComfyUI template, in the hosted API and as a starting point in the alternative generator above.

Where it shines and where it stumbles

Community feedback lines up with the design trade-offs. Speed and realism get the praise; variety gets the complaints. The team reported eighth place overall and first among open models on the Artificial Analysis text-to-image leaderboard on December 8, 2025. That ranking is dated: by October 2026 the model was no longer in that board's top 20 as newer releases arrived. Gigazine's hands-on review noticed repeated compositions across seeds, look-alike faces in group scenes and trouble with Japanese text, while the slower base model did better on illustration and range.

Adoption signals in October 2026: about 12,100 GitHub stars, about 1.07 million 30-day downloads of Unsloth's GGUF builds and about 241,000 for one popular LoRA pack. Downloads show reach, not quality.

What people actually make with it

Because a render takes seconds, people use the model for volume work where they would otherwise give up after a few tries. Sellers on Chinese marketplaces like it for banners that mix both languages, and short-drama and game teams use it for fast mood boards before a final render elsewhere.

• Portrait and fashion series: dozens of outfit and lighting variations of the same adult model in one sitting.
• Bilingual e-commerce art: product scenes with Chinese and English headlines for shop pages and social feeds.
• Pre-visualization: key frames for short dramas, ads and game pitches.
• Character LoRAs: community add-ons for specific looks, costumes and art styles.
• Pose matching: the ControlNet Union model copies a stance or layout from a sketch.
• Older hardware: GGUF builds made it a common pick for laptops and mid-range gaming cards.

No GPU? The hosted z-image-turbo endpoint

Alibaba Cloud Model Studio serves the model under the ID z-image-turbo, and it is the only member of the family offered there. Each request carries one user message with a single text prompt of up to 800 characters, plus optional size (512×512 to 2048×2048 total pixels, default 1024×1536), seed and a prompt_extend flag that lets a language model rewrite the prompt and return its reasoning. You get one PNG per call, its link expires after 24 hours, and the limit is 120 requests per minute. Prompts and outputs pass a safety check, and batch jobs, fine-tuning and function calling are not available, which makes the endpoint convenient but strictly moderated.

Up against FLUX.2 and Qwen-Image

The technical report pitches the series as a 6B answer to open models of 20B and more, such as Qwen-Image, HunyuanImage 3.0 and FLUX.2. Treat that as the authors' framing. In practice the gap is running cost: a small model with 8 steps is the easiest of the three to keep on one consumer card, while the larger ones may cope better with crowded scenes. Test with your own prompts before switching.

Sizes and licenses follow the technical report; check each model card before relying on them.

Trying the uncensored route in a browser

The browser path skips downloads and drivers. It does not run this model, so the look will differ; Spicy AI uses its own image models.

• 1. Press the button at the top of the page to open the image generator.
• 2. Paste a template from this page and describe every person as an adult.
• 3. Choose an aspect ratio and create up to four images per run.
• 4. Change only one thing per retry: light, pose or outfit.
• 5. Upload a photo you own to switch to image-to-image editing.

Five habits that save failed renders

1. Leave CFG at 1 and steps at 8; move to the base model when you need a negative prompt.
2. Download the text encoder and VAE alongside any GGUF file.
3. Vary the scene in words, not only the seed, to break repeated layouts.
4. Put exact on-image words in quotes and proofread small text before publishing.
5. Describe every character as an adult (18+) and never use a real person's likeness without consent.

Sources

• GitHub repository (Tongyi-MAI)
• Turbo model card (Hugging Face)
• Official project blog
• Technical report (arXiv 2511.22699)
• ComfyUI repack (Comfy-Org)
• GGUF builds (Unsloth, third party)

More uncensored AI Image models

• uncensored z image
• uncensored qwen image
• Uncensored Midjourney
• Uncensored DALL·E
• Uncensored Stable Diffusion
• Uncensored Flux
• Uncensored Nano Banana
• Uncensored Nano Banana 2
• Uncensored Nano Banana 2.1
• Uncensored Ideogram
• Uncensored Leonardo AI
• Uncensored Seedream
• Uncensored Seedream 5.0
• Uncensored Krea
• Uncensored Kroma
• Uncensored Qwen-Image
• Uncensored Adobe Firefly
• Uncensored GPT Image 2.5
• Uncensored Text to Image
• Uncensored Z-Image
• Uncensored FLUX.2
• Uncensored FLUX.2 [dev]
• Uncensored FLUX.2 [pro]
• Uncensored GPT Image 1.5