Is GPT Image 2 Uncensored? No, and Nano Banana Stops at Advanced NSFW. What Each One Refuses

Short answer: no. Based on our September 2026 tests, GPT Image 2, 2.5 Sunburst and 2.5 Flare refused advanced NSFW imagery, black lace lingerie and a clothed intimate couple. Nano Banana 2 and Pro drew the lingerie and the couple, then stopped at advanced NSFW imagery. "Advanced NSFW imagery" here means a direct L3 request. All five are labelled Filtered: OpenAI's and Google's own safety systems do the refusing. For adult work, choose a model with tested support for the content you need. There is no switch that removes these restrictions, and rewording is not a dependable unlock.

What each model refuses

Probes: advanced NSFW imagery, black lace lingerie, and two clothed adults in an intimate moment; for editing, the same three kinds of change to a clothed portrait (the third being an open jacket over a camisole). ✅ came back as asked · ❌ refused.

• GPT Image 2 refuses from lingerie up, and once refused a clothed red satin gown prompt.
• GPT Image 2.5 Sunburst and Flare both refused every probe, making both stricter than Image 2 on this editing set. Sunburst even refused opening a jacket to show a camisole.
• GPT Image 2.5 Flare refused every probe and, once, a clothed neon poster prompt.
• Nano Banana 2 and Pro refuse at advanced NSFW imagery, not before: lingerie and a clothed embrace render as asked. The line is a direct L3 request. Partial exposure in an artistic portrait varied: Nano Banana 2 rendered it, Pro returned a toned-down version (billed as a normal result) and all three GPT Image models refused. Editing is stricter than generating: both refused an edit that put lace lingerie on a clothed portrait.

GPT Image: no setting unlocks it

GPT Image 2 has an OpenAI moderation control, and requests through SpicyAPI already go out at its most permissive level, so there is no field to change. Refused GPT requests are not billed.

Keep it for what it is built for: readable text, posters and covers, exact edits, product shots and clothed editorial work.

Nano Banana: lingerie yes, advanced NSFW no, and a billing trap

Google's pair is more permissive, but the two refuse differently, and that changes your bill:

• Nano Banana Pro returns a failure ("filtered out"). Refused requests are not charged.
• Nano Banana 2 returns a completed task carrying one sentence of text and no image, and bills the full price. Its model page states the rule: a refusal that comes back as a short text reply counts as a completed generation; other refusals are not charged.

So with Nano Banana 2, completed does not mean you got an image. Check the result contains one, and budget for the odd text-only reply at volume.

Clothed fashion and product work can be refused too

Both makers check the finished image, not only your prompt, so a harmless prompt can still be refused.

• GPT Image: developers report output-stage adult-content refusals even at moderation: low, on prompts as plain as a woman getting ready for a party. Naming the outfit (for example summer clothes) helped in one case.
• Nano Banana: developers report IMAGE_SAFETY blocks on underwear catalogue shots and on swimwear even with every safety category set to BLOCK_NONE. Behaviour varies between prompts and over time.

For lingerie, swimwear or underwear catalogue work, test your prompts on the Filtered models and keep a fallback. If a brief regularly needs skin, use an Uncensored model from the start rather than paying for refusals on Nano Banana 2.

Prices

Starting prices per image at the stated settings:

What to use for adult content

Each model below is labelled Uncensored: no content filter of ours sits in front of it.

For a wider comparison, see our ranking of NSFW image generators.

In short: keep GPT Image and Nano Banana for text in images, product shots, clean edits and clothed editorial work, and send adult briefs to an Uncensored model instead of rewording for a Filtered one.

Test your prompt in the Playground first

Check a brief on the models before you write any code. A refused GPT Image request is not billed, so the comparison costs nothing when it refuses.

1. Create an account, then open Qwen Image 2.1. The page opens on the Playground tab in Form view.
2. Paste your brief into Prompt. Set Resolution to 1k, the cheapest tier, and press Run model. You need a funded account to run.
3. Open GPT Image 2, paste the same prompt, set Quality to low for a draft and run it. For an Uncensored alternative with a different look, open Seedream 5.0 Pro the same way.

Spicy Studio does the same without a form: open Studio, choose Image, type the prompt and pick the model in the model picker.

Then run it from code

For this minimal, prompt-only body, moving to Qwen is a one-line change to model. In an existing request, also remove or replace fields the target schema does not accept, such as GPT-specific quality or resolution values.

curl -X POST https://api.spicyapi.ai/api/v1/jobs/createTask \
  -H "Authorization: Bearer $SPICY_API_KEY" \
  -H "Content-Type: application/json" \
  --data '{
    "model": "alibaba/qwen-image-2.1/text-to-image",
    "input": {
      "prompt": "Editorial fashion photograph of an adult woman in black lace lingerie on rumpled white sheets, soft window light, 35mm film grain"
    }
  }'

Waiting for the result, downloading it, errors and retries are covered in the tasks guide, and the image generation guide lists the image fields.

Or let an Agent run it

When an assistant should run the generations, all three official tools show the price and ask you before a paid call.

• CLI. npx @spicyapi/cli tasks create --model alibaba/qwen-image-2.1/text-to-image --input-json '{"prompt":"Editorial fashion photograph of an adult woman in black lace lingerie on rumpled white sheets, soft window light"}' --wait (set SPICY_API_KEY first). See the CLI docs.
• MCP server. Add the MCP server to your assistant, then ask: "Run this prompt on Qwen Image 2.1." It calls spicyapi_task_create, which stops for your price confirmation, then spicyapi_task_wait.
• Skill. npx skills add spicyapi-ai/spicy-skill installs the Agent Skill, which makes a coding assistant read the live model schema instead of guessing fields.

FAQ

Is GPT Image 2 uncensored?

No. Its label is Filtered. Sunburst and Flare both refused every probe in our editing set, including the clothed jacket edit Image 2 allowed.

Does moderation=low turn the filter off?

No. Requests already use the most permissive level.

Is Nano Banana uncensored?

No, but it goes further than GPT Image: lace lingerie and a clothed intimate couple render, advanced NSFW imagery does not.

Am I charged when a request is refused?

Not on GPT Image or Nano Banana Pro. On Nano Banana 2, a text-only reply on a completed task is billed; a plain failure is not.

Can I get around the refusals with different wording?

There is no wording-based switch that disables the makers’ restrictions. For normal clothed fashion or product work misclassified as unsafe, specifying the outfit or clarifying the non-graphic brief sometimes helps, as the developer reports above show. That is not a reliable way to unlock restricted adult content.

Your part

SpicyAPI adds no platform-level content filter; the model and upstream service still apply their own rules. Everyone in an image must be an adult. Don't depict a real, identifiable person without their consent. You are responsible for the law where you publish.