Is Z-Image Turbo censored?
Z-Image Turbo's Apache 2.0 model card mentions no safety filter. Refusals come from hosts: fal's checker is on by default and blacks out flagged images.
Z-Image Turbo is released as open weights by Alibaba's Tongyi-MAI lab under the Apache 2.0 licence, and its Hugging Face model card doesn't mention a safety filter. When a prompt is refused or an image comes back black, that is the host's checker, not the model. fal, for example, runs one by default. Versely's catalog row lists no filter either way.
"Is Z-Image Turbo censored?" is two questions glued together. One is what the model and its card say. Tongyi-MAI's model card, read on 16 October 2026, describes Turbo as the distilled member of a 6B-parameter family that runs in 8 steps and "fits comfortably within 16G VRAM consumer devices". It has no section on content filtering. The other question is whether a host runs a safety checker on prompts or outputs. Hosts often do. A checker is a gate on a door. It is not a field on the model card, and it is not a public blacklist you should expect us to reprint.
The published slug is z-image-turbo. This page will not invent a list of blocked topics. If a generation is refused, that refusal is a host or vendor policy in effect at run time, not a feature named in the Versely description.
What the catalog actually says
Z Image Turbo is an efficiency-tuned text-to-image model built for speed. Single-credit generations for drafts, mood boards, and rapid A/B. 1 credit. Marked fast. 4K max output. Qualities SD, HD, and 4K. Aspects 1:1, 16:9, 9:16, 4:3, 3:4. No audio. No duration. Category: text-to-image. Released in the catalog data as 27 November 2025. Features listed: Turbo, Fast, Quality. Styles listed: Realistic, Artistic, Anime. There is no "safety," "NSFW," "censored," or "uncensored" flag on that card.
That absence is the honest catalog answer. We do not add a filter list to fill the gap. We do not claim the row is uncensored. We do not claim it will draw anything you type. Third-party write-ups disagree with each other on purpose: local ComfyUI users talk about workflow-level filters they did or did not load; hosted demos talk about a checker they bolted on; fine-tunes advertised as "NSFW unlock" exist because someone wanted different behaviour than the stock hosted demo. None of those pages are this catalog.
A safety checker, in the sense Hugging Face demos have used for years, is often a separate small model (the old CompVis Stable Diffusion checker is the famous one) that inspects the output or the prompt and replaces or blocks a result. It is not Z-Image Turbo's architecture. It is furniture. Turn it off locally and you have a different stack than the demo space. Leave it on and you will call the model censored when the space refused. That confusion is the query.
One host's switch. fal's Z-Image Turbo API has an enable_safety_checker setting, on by default. Its docs say: "Disabling it requires account authorization; unauthorized requests are always checked, and images flagged as unsafe are returned as black images." The same response carries a has_nsfw_concepts flag for each image. So on that host a black square is the checker at work, and the weights underneath are the same Apache 2.0 checkpoint.
Tongyi-MAI's Z-Image family is a weights story: Turbo and the Z-Image foundation model are published, and the card lists Z-Image-Omni-Base and Z-Image-Edit as "To be released" on 16 October 2026. Host policy is a product story. Versely's product story for this slug is: 1-credit stills, 4K ceiling, lock the still before you buy motion. The sibling is Z Image Turbo: lock this still before you buy motion. Censorship is not that sibling's topic either.
Checker versus "censored" as a personality
Creators use "censored" to mean three different failures.
The prompt never ran. A host blocked the request before diffusion. That is a policy gate. You will not get a picture. Trying a "spelling trick" is how people get accounts banned on platforms that log it, and this page offers no bypasses.
The prompt ran and the picture came back blurred, black, or replaced. That is often an output checker. Same gate, later in the pipe. The weights may have produced something. You were not shown it.
The picture ran and looks prudish compared to a local fine-tune. That is training data, alignment, and your comparison set. A Turbo distill trained for photoreal product stills will not match a fine-tune advertised as an unlock. It is also not proof of a published Versely filter. It is a different checkpoint.
None of those three is a reason to skip Turbo for a serum bottle, a kitchen, a logo draft, or a 9:16 thumbnail. The row is a mood-board machine. Rapid A/B at 1 credit is the described job. Ten looks, one winner, then stop. Animate the winner on image-to-video. If one object is wrong, fix it on the locked still with a one-line instruction in Qwen Image Edit 2511 rather than hunting for a "less censored" model.
Illegal content is not a checker debate. Child sexual abuse material, non-consensual intimate imagery, and the rest of the criminal category are banned whether or not a slider exists. This page will not itemise a grey list around that line. If you are asking "is it censored" because you want a pornography model, you are asking for a different product than a 1-credit packshot draft row. We will not document how to obtain one.
Brand safety for ads is a fourth, ordinary concern. A checker that blocks a swimsuit is not the same as a checker that blocks a crime. If a commercial still refuses, simplify the prompt: product, set, light, no celebrity likeness, no living-person impersonation you do not have rights to. Likeness and trademark are legal facts, not "censorship of art."
What to do instead of hunting a filter list
Run the stills job you actually have.
Product, editorial, thumbnail, mood board: Turbo, 1 credit, pick SD/HD while sketching, 4K when the still might be animated or printed. Aspects on the card. The text-to-image tool is the door.
If a prompt refuses, drop the celebrity, the sexual clause, the violence, or the private photo you do not own, and run the commercial still. If it still refuses, try a different image row that is built for the job, not a "jailbreak" of this one. Record the refusal in your notes if you version assets: slug, prompt, "blocked at host." That is AI asset versioning, not a bug file we can reverse from a blog post.
If you are comparing local versus hosted: local weights plus your own workflow can omit a checker that a space includes. That does not make the Versely card "uncensored." It makes two stacks. We serve a hosted row. We do not publish its hidden classifier labels.
If you need photoreal people for a UGC-style ad, you still have likeness, disclosure, and Article 50 (from 2 August 2026) when the clip is a realistic synthetic of a real person for EU audiences. A checker is not that label. The label is your overlay.
The text-to-image arena rank in the catalog (Elo 935, #70 on 16 October 2026) is not a morality score either. It is a preference rank. It does not mean "more censored than GPT Image 2."
The test
Are you generating a packshot, a kitchen, a type-heavy thumbnail? Censorship is the wrong question. Use Turbo. Lock the still.
Are you asking for a published list of blocked tags on this slug? There isn't one on the card. We will not invent it.
Are you asking whether some host somewhere filters Z-Image Turbo? Yes, some do. That sentence will still be true next month, and the list will still not belong in this catalog.
Use the row for speed. Treat refusals as policy, not as a missing spec. Move the job to a row that matches a legal commercial still if this one will not run it.
Draft the stills pack on Z Image Turbo, then animate the keeper. Sign up when the job is a 1-credit mood board, not a scavenger hunt for an unlock.
FAQ
Does Versely list Z-Image Turbo as uncensored?
No. The published card lists speed, 1 credit, 4K stills, and the usual image knobs. It does not list censored or uncensored, so neither can be read from the absence.
Why did a prompt fail if there is no filter list?
A host-side safety checker or vendor policy can still refuse a run. That gate is not printed on the model card. A refusal is not a catalog bug, and no blog post can name every blocked class.
Is a local "NSFW unlock" fine-tune the same model?
No. Fine-tunes are different weights. The Versely slug is the stock Turbo stills row. A community unlock is another checkpoint with another policy problem. We do not serve it under this slug.