Is Ideogram Uncensored? Ideogram v4.5 NSFW Test Results

Short answer: no. On 5 October 2026 we ran Ideogram v4.5 through our content ladder (L1 to L4) and it refused 11 of 13 requests. The only two it delivered were a bikini beach scene and a neutral lingerie catalogue photo. Refusals came back in about 2 seconds, before any image was made, with the message "Content flagged as potentially sensitive. Please try different prompts or images."

If you need an image model that does not refuse adult-oriented requests, Ideogram v4.5 is not it, at least on the route we tested. The rest of this page shows the full ladder, what the filter seems to be doing, what Ideogram is still positioned for, and which models on SpicyAPI rendered the same ladder as written.

At a glance

How we tested

We used the same ladder we use for every image model. Five levels, each stronger than the one before, with a neutral gloss for each:

We did not run L5 on Ideogram. Once L2 through L4 came back refused, there was no reason to push further up the ladder.

Ideogram v4.5 was reached through a third-party hosted API. A hosted route can add its own filter on top of the model maker's, and from the outside the two look identical; we come back to this below.

The full ladder: what was delivered and what was refused

Two requests passed and the rest of the ladder did not. This is the whole table for the 13 requests, which include the runs where we turned prompt expansion on and off.

Counting repeats and the prompt-expansion comparison, that is 11 refusals out of 13 requests. The two that passed share a pattern: both describe clothing as clothing, with nothing in the wording that points past it.

What the delivered image looks like

The bikini image is a perfectly ordinary photo, and it is a fair picture of what Ideogram v4.5 will draw. It shows a woman in a red two-piece swimsuit standing at the edge of the surf at dusk, hair wet, palm leaves in the corner of the frame. Nothing in it went past what you would see in a travel brochure.

The bikini beach scene Ideogram v4.5 delivered in our run on 5 October 2026, text to image, 1k. This is the strongest request it fulfilled.

The model can draw; the route in front of it decides what it may draw, and in our run that line sat at swimwear.

The refusal: message, timing and cost

Every refusal looked the same. The message was always "Content flagged as potentially sensitive. Please try different prompts or images." It arrived in about 2 seconds, which is far too fast for an image to have been generated and then checked. In other words the request was judged before any generation started.

Two practical consequences follow:

• Refused requests were not charged. A refused call costs you a couple of seconds, nothing else.
• The message gives you no hint about what to change. It does not say which phrase triggered it, so "try different prompts" is the only advice on offer.

The filter is semantic, not a keyword list

Rewording did not get past it. One of our requests was written without any word a keyword list would normally flag, using a euphemism for the same scene. It was refused as well, and with the same message.

That is the main finding behind the headline number. A keyword blocklist would have let the euphemistic prompt through. Because it did not pass, the check is reading what the prompt means, not scanning for particular words. For you as a developer this means that prompt engineering around the filter is not a dependable plan. It also means the filter can catch requests that read as innocent to a human.

We also ran prompt expansion both on and off, in case the model's own rewriting step was what tripped the check. It made no difference to the outcome.

Whose filter is it?

We cannot tell, and we would rather say so than guess. We reached Ideogram v4.5 through a third-party hosted API. The refusal could be part of Ideogram's own API, or something the host puts in front of the model, or both. Nothing in the response tells them apart.

So the accurate reading of our result is narrow: Ideogram v4.5, on the route we used on 5 October 2026, refused 11 of 13 requests. It is not a statement about every way of reaching Ideogram, and we did not test any other route.

What Ideogram is still good for

Ideogram positions itself around typography and poster work, and that is where it is meant to be used. We did not test text rendering, layout or poster quality in this run, so we make no claim about how well it does there. The point for this article is simpler: if the job is a poster, a logo mock-up or a title card, the content filter is unlikely to matter and the publisher's own positioning is the reason to look at it.

If your job is adult-oriented imagery, a model that refuses at swimwear is the wrong tool whatever else it is good at.

For the same test on FLUX 3 through the same kind of route, and for how a differently built model behaves, see our other level-by-level write-ups: Is FLUX uncensored? covers FLUX.1 Dev LoRA, and Is Z-Image uncensored? covers the Z-Image family.

Uncensored alternatives on SpicyAPI

These models rendered our content ladder as written. On SpicyAPI the platform adds no content filter in front of them. That is not the same as every image being perfect: anatomy and hands still need a seed reroll now and then, like with any image model.

Anima Turbo: the illustration route

Anima Turbo is the one to start with if illustration is what you need. It is a 2 billion parameter anime and illustration model from CircleStone Labs, and it accepts Danbooru-style tags, plain sentences, or a mix. In our tests it took 25 requests across L1 to L4, text to image and image to image, with and without LoRAs, female and male subjects, and produced zero refusals and zero substitutions. L4 came back as written on the first try.

That is the opposite of the Ideogram result above, using the same ladder. Our Anima Turbo write-up has the full test.

Anima Turbo, text to image, 3:4, 1k, a clothed suggestive scene with the sensitive rating tag.

The rating tags set the tone, because there is no negative prompt. The model knows safe, sensitive and nsfw, plus one stronger tag; the publisher documents the last two as the ones that remove restraint. We only print prompts with the first two. With a clothed description, the two gave almost the same picture:

The same lounge prompt and the same seed, safe on the left and sensitive on the right.

A prompt that works as a starting point looks like this:

masterpiece, best quality, score_7, sensitive, 1girl, solo, adult woman, mature female,
long brown hair, white bathrobe slipping off one shoulder, standing by a large window at dawn,
holding a coffee cup, city skyline, soft light, calm smile

Every character is an adult, and the prompt says so. Run it with aspect_ratio set to 3:4 and resolution set to 1k.

Men, couples and every pairing

The model does not default to one kind of subject. Our tests covered female and male subjects, and the pool below includes a man in a doorway, a couple dancing close and two women on a sofa. All adults, all clothed.

Anima Turbo, text to image, 3:4: a mature man in an open shirt in a doorway.

Anima Turbo, text to image, 16:9: two adults dancing close in a dim bar.

Anima Turbo, text to image, 16:9: two adult women in silk pajamas on a sofa.

Keeping one character across scenes

Fixing the seed and changing the scene keeps a character recognisable. The three frames below use the same character description (silver bob, golden eyes, a beauty mark) in three outfits and places, with the same seed. For a series or a comic, that consistency matters more than any single frame.

Anima Turbo, text to image, 3:4, same seed, three prompts: turtleneck in a gallery, red dress on a balcony, oversized shirt on a windowsill.

Anima Turbo LoRA: styles you bring yourself

The LoRA family adds up to three adapters of your own to the same model. A style LoRA changes the look completely without a different base. Two we used, both public and permissively licensed: a PC-98 era game look and a monochrome manga ink look.

Anima Turbo LoRA, text to image, 4:3, 1k, PC-98 style LoRA at scale 1.

Anima Turbo LoRA, text to image, 3:4, monochrome ink LoRA at scale 1.

Two facts from our tests: a LoRA trained for a different base model, such as a FLUX one, loads without an error and changes nothing; and about 0.5 to 1.2 is the useful scale range, while 4 broke the image into abstract shapes. Our Anima LoRA style guide covers the details, and photo to anime with Anima Turbo shows image to image at different strengths.

When you need a photoreal base

For photoreal work, the options are Z-Image Turbo LoRA and Qwen Image 2512 LoRA. Anima does not do photorealism; it is trained on anime and illustration data, so its outputs read as illustration and never as photographs of real people. That is a feature for some uses and a limit for others.

• Z-Image Turbo LoRA is photoreal by default and turns anime when you load an anime LoRA. L3 and L4 rendered in our text-to-image tests.
• Qwen Image 2512 LoRA takes your LoRAs and also edits existing images. L3 and L4 rendered in our tests.

Both load LoRAs you supply, and loading them adds no separate charge. Check the live price on each model page, since prices are shown there and not here. We do not compare their image quality here beyond what our content tests recorded.

A note on responsible use

Adults only, and you are responsible for what you make. Every character in every example here is an adult, and every prompt says so. Do not generate or share images of real people or their likenesses without their consent. Keep your own outputs private unless you have the right to publish them. SpicyAPI does not write adult prompts for you, and the way you use these models is your responsibility.

FAQ

Is Ideogram v4.5 uncensored?

No, not in our test. On 5 October 2026 it refused 11 of 13 requests across our content ladder, including L1 with flirtatious wording, L2, L3 and L4. The two requests it delivered were a bikini beach scene and a neutral lingerie catalogue photo.

Does turning prompt expansion off change the result?

No. We ran it both on and off and the outcome did not change. The refusals came from a check that runs before generation, not from the prompt rewriting step.

Does Ideogram charge for refused requests?

Not in our run. Refused requests came back in about 2 seconds with no image and were not charged.

Can I get past the filter by rewording the prompt?

It did not work for us. A euphemistic prompt with no flagged words was refused with the same message, which tells us the filter reads meaning rather than matching a word list. We would not build a workflow on rewording around it.

Is the filter Ideogram's or the host's?

We cannot tell. We tested Ideogram v4.5 through a third-party hosted API, and the refusal could come from Ideogram's own API, from the host, or from both. Our result describes that route on that date only.

What is Ideogram good for, then?

Ideogram positions itself around typography and posters. We did not test those, so we do not rate it there. If your work is posters or title cards, the filter is unlikely to get in your way.

Which SpicyAPI models rendered the ladder as written?

Anima Turbo, Anima Turbo LoRA, Z-Image Turbo LoRA and Qwen Image 2512 LoRA all rendered L3 and L4 as written in our tests. Anima Turbo had zero refusals and zero substitutions across 25 requests. FLUX.1 Dev LoRA rendered L3 but softened L4 in our tests.

Is Anima Turbo suitable for realistic images?

No. Anima is an anime and illustration model and does not do photorealism. For a photoreal base, use Z-Image Turbo LoRA or Qwen Image 2512 LoRA.