Is FLUX 3 Uncensored? We Tested 12 Prompts (It Blocks Anything Past Lingerie)

No. In our test FLUX 3 is not an uncensored model: it drew a bikini beach scene and a neutral lingerie catalogue photo, and refused 10 of our 12 requests, every one of them past that point. The refusals came back in about two seconds, before any image was made, and were not charged. A prompt with no flagged words in it was refused too, so the filter reads meaning, not a word list.

We ran the test on 5 October 2026 through a third-party hosted API, text to image, 1k. We cannot tell whether the filter belongs to Black Forest Labs' own API or to the host we used, so read the result as "FLUX 3 on that route", not as a verdict on the weights. If you need an NSFW-capable image model, the last sections list four we have run up the same ladder, and it is a short list of models that did not refuse.

At a glance

The ladder: what we asked for and what came back

FLUX 3 stopped at the first request that was meant to be suggestive. We use a five-level ladder for every image model we test, from L1 to L5, and we add a few side probes to see whether wording or packaging changes the answer. Here is every request in this test.

Some rows were run more than once, which is how 12 requests fit into nine rows. The result of each row was the same every time we repeated it.

Two things stand out. The first is how low the line sits. The refused L1 prompt and the delivered L1 prompt are both lingerie; the only difference is the wording of the request. The second is that none of the side probes found a way through. Wrapping a request as a classical painting study did not help, a male subject was treated like a female one, and indirect phrasing was caught as well.

The one image we can show from this test is the delivered swimwear scene. It is a photographic rendering, and everything past it was refused.

What the refusal looks like

A refused request fails fast, with one fixed message and no image. The task ends as failed with an empty result and this text:

Content flagged as potentially sensitive. Please try different prompts or images.

In our runs it came back in about 2 seconds, which tells you the check happens before generation, not after. Refused requests were not charged. None of the request fields we saw has a safety or strictness setting, so there is nothing to switch off: the prompt, the aspect ratio, the resolution, the output format and a prompt-expansion toggle are all there is.

One practical effect of an upfront check is that you cannot tell what tripped it. The message names no word and no category. That is why creators who hit it tend to rewrite the prompt, which brings us to the next section.

Why a euphemistic prompt still failed

Because it was refused too, the filter is semantic: it judges what the prompt means, not which words it contains. We wrote one request with none of the usual flagged words in it and described the scene indirectly. A keyword list would have let it through. FLUX 3 refused it in the same two seconds as the blunt ones.

The same pattern shows up inside the lingerie rung. Two prompts about the same garment got opposite answers, and the thing that changed was the implied intent, not a single banned term. In practice this means three common workarounds do not help on this route:

• Softer vocabulary. The meaning is still read.
• Art or fashion framing. Our classical painting study was refused.
• Switching the subject. The male subject was refused like the female one.

The useful reading is that the filter is doing a job on purpose. It is a content classifier sitting in front of the model, and we cannot tell whether it was put there by the model's maker or by the host. If the filter belongs to the host, the same weights could behave differently elsewhere. If it belongs to the maker, it will follow the model wherever its official API is used. We have no evidence either way and do not want to guess.

What this means for NSFW creators

If your work lives anywhere past lingerie, FLUX 3 on this route will stop you on request one, so pick the model before you plan the pipeline. The costs of finding that out late are the real problem, not the two-second refusal itself.

• Pipelines break at the first refused call. A batch job that mixes safe and mature prompts will fail on the second kind. Handle the failed status in your code, and do not retry the same prompt: the answer will not change.
• You cannot tune around it. There is no setting, no negative prompt and no rating tag that lowers the bar, so prompt engineering is wasted time on this route.
• It is still a useful model for the work it does accept. Swimwear, fashion and catalogue imagery went through. If a project is mostly that, the refusals are a boundary, not a defect.
• Photographic output is part of the picture. FLUX 3 produced a photographic woman. Many creators who work with adult themes prefer illustration, partly because the result reads clearly as fiction. The models below are stronger fits for that.

Our older test of the previous generation reached a different, softer result: that model softened some levels without refusing and ran with no filter at all in front of it. That article is about FLUX.1 Dev with LoRAs, which is a different model; nothing in this test changes it.

The alternatives that rendered the same ladder

Four models on SpicyAPI rendered their ladder levels as written in our tests, with no content filter in front of them. The table says what each one is and what we saw. We have not compared their image quality head to head here, so treat the last column as a content result only.

"Rendered" means the image matched the request as written, not a substitute. For FLUX.1 Dev LoRA we keep the caveat in the table on purpose: if you like the FLUX look and want a filter-free route, it works up to L3, and L4 came back softened. The longer write-up is in our FLUX.1 article, and the Z-Image test covers the other photoreal option.

Anima Turbo: the one we ran hardest

Anima Turbo is the model we put the most requests through on the same day as the FLUX 3 test. In 25 requests across L1 to L4, text to image and image to image, with and without LoRAs, female and male subjects, and two seeds for L3 and L4, we saw zero refusals and zero substitutions. L4 rendered as written on the first try. A separate run of 60 aspect-ratio requests at an L2-level prompt delivered all 60.

It is an anime and illustration model, not a photographic one. For adult-themed work that is a practical advantage: the outputs read as illustration, never as photographs of real people. It also means it will not give you a photoreal look, which FLUX 3 does well for the content it accepts.

The prompt for that image is an ordinary clothed scene. It uses the sensitive rating tag and says plainly that both people are adults:

masterpiece, best quality, score_7, sensitive, 1girl, 1boy, hetero, adult couple, mature female, mature male, slow dancing close together in a dim jazz bar, she wears a green satin dress, he wears a dark suit with loosened tie, foreheads touching, his hand on her waist, warm lamp light, intimate mood

Same ladder, different subjects

Our FLUX 3 refusals did not depend on the subject, and neither did the Anima results. We ran female and male subjects through the same levels and none was refused. Here are two more clothed examples from the same session, both prompts tagged sensitive and both with adult characters.

masterpiece, best quality, score_7, sensitive, 2boys, adult men, mature male, couple on a rooftop at night, one with short blond hair in a leather jacket, the other with black hair in a white t-shirt, arm around shoulder, leaning close, city lights, warm smile

The rating tags are the dial

On Anima, the rating tag in the prompt sets the tone, because there is no negative prompt. The model knows safe, sensitive and nsfw, plus one stronger tag we do not name. The publisher documents nsfw and the stronger tag as the ones that remove restraint. In our test, with a clothed description, safe and sensitive gave almost the same picture at the same seed. Every prompt we print starts with safe or sensitive, and we do not print prompts that use the stronger tags.

Anima Turbo LoRA: your own style, same ladder

Anima Turbo LoRA runs the same model with up to three LoRAs of your own, each a direct HTTPS link to a .safetensors file with a scale from 0 to 4. L3 and L4 rendered with it too. LoRAs trained for another base model load without an error and change nothing, so check the model card says Anima. If you are new to adapters, our LoRA style guide for Anima shows what a scale does.

A fixed character description can also hold steady across outfits and places.

Where FLUX 3 is the right call, and where it is not

Choose by the content you need, not by the model's reputation. The table is a decision aid, built only from what we ran.

Where the table says "not tested for this", we did not run that case and are not guessing. The honest summary: FLUX 3 is a fine fit for work that stays at or below lingerie, and a poor fit for everything above it.

Your part

Everything on this page assumes adults. Every character in our prompts is an adult, and yours must be too. Do not generate real, identifiable people or their likenesses without consent. Keep your outputs private and do not publish anything you would not want attached to your name. You are responsible for what you generate and for where you publish it. We do not review prompts for you.

FAQ

Is FLUX 3 uncensored?

No, not in our test. Through a third-party hosted API on 5 October 2026, FLUX 3 delivered a bikini beach scene and a neutral lingerie catalogue photo and refused 10 of 12 requests, starting with lingerie written in a flirtatious way. We cannot tell whether the filter belongs to Black Forest Labs' API or to the host.

How fast does FLUX 3 refuse a prompt?

About two seconds, before any image is generated. The message is "Content flagged as potentially sensitive. Please try different prompts or images." Refused requests were not charged in our runs.

Can I get around the FLUX 3 filter with different wording?

Not in our test. A prompt with no flagged words was refused, so the filter reads meaning, not keywords. Art-style framing and a male subject were refused too, and the request fields have no safety setting to change.

Is FLUX 3 the same as FLUX.1 Dev?

No. They are different models with different results. Our FLUX.1 Dev tests, including its LoRA route, are in the FLUX.1 article. On our catalogue's content tests FLUX.1 Dev LoRA rendered L3 and softened L4.

Which image model should I use for NSFW work instead?

It depends on the look. For anime and illustration, Anima Turbo and Anima Turbo LoRA rendered L3 and L4 as written in our tests. For a photoreal look with your own LoRAs, Z-Image Turbo LoRA rendered L3 and L4 and FLUX.1 Dev LoRA rendered L3. Our roundup of NSFW image generators puts more models side by side.

Does Anima Turbo really never refuse?

In our 25 ladder requests and 60 aspect-ratio runs, we saw zero refusals and zero substitutions. That is a result from one day of testing, not a promise about every prompt. Anatomy and hands can still need a seed reroll, like any image model. Read the Anima Turbo test for prompting and limits.

Try the alternatives

Open Anima Turbo in the Playground, paste the dance prompt above and change the seed until the picture is right. The live estimate on the page shows the cost before you submit.