Moderated vs private AI image models — which to pick on eroq

Every image model on eroq sits behind the same prompt box, but they do not all play by the same rules. Some are moderated, and your prompt may be retained under the engine's terms. Others are fully private: eroq runs them itself, end to end. Which family you pick decides three things before you type a word: who can see your prompt, whether the model can read your pictures, and what it will agree to render. Here is the map of all 13 image models, and a pick for each common situation.

The two families

	Moderated	Fully private
Models	GPT Image 2.5 Sunburst and GPT Image 2.5 Flare (OpenAI), Nano Banana 2.1 and Nano Banana Pro (Google), Seedream 5.0 Pro (ByteDance), Flux 3 Image (Black Forest Labs), Grok Imagine 2.0 (xAI), plus eroq's Flash Image and Flux Image	Krea2, Eroq Uncensored, Eroq One, Anima
Your prompt	may be retained under the engine's terms, never used for training	eroq runs the engine itself: fully private
With Private mode	the result is never stored	fully compatible: nothing of the render is stored
Reference pictures	none: text-to-image only	Eroq One up to 10, Krea2 and Eroq Uncensored up to 4, Anima none
Editing	none	Eroq One by instruction, Krea2 and Eroq Uncensored in place
Content	moderated, each engine under its own rules	Eroq Uncensored, Eroq One and Anima uncensored within the acceptable-use policy; Krea2 clean photoreal
Credits per image	10 (Flash Image) to 124 (Nano Banana Pro)	10 (Krea2) to 10 (Eroq One)

Both families share the basics: five formats at one price, one to four takes per run, automatic refunds on failed renders, and the same POST /v1/images/generations call in the API. Access is where they part: the private engines, Flash Image and Flux Image run on every account, pay-as-you-go included, while Flux 3 Image and Grok Imagine 2.0 open on Hobby and the GPT Image, Nano Banana and Seedream models on Creator.

Privacy: what "may retain" actually means

When you render on GPT Image, Nano Banana, Seedream, Flux 3 or Grok Imagine, your prompt travels to that model's maker, because the maker's servers draw the picture. It may keep the prompt under its own terms. It does not use it for training. Flash Image and Flux Image follow the same retention rule.

Private mode changes eroq's side of the story completely. Turn it on in the studio (or send private: true through the API) and eroq writes nothing of the render to storage: no file, no Library entry, no prompt on the job. The picture goes straight to your browser. The usage ledger still records the charge, with your prompt replaced by stars. What Private mode cannot do is stop a moderated engine from keeping the prompt, so on those models the honest description is "result never stored, prompt possibly retained upstream".

On the private engines (Krea2, Eroq Uncensored, Eroq One and Anima) nothing leaves eroq: eroq runs them itself, end to end. The model picker says which is which, with a « Private mode ✓ » badge on the engines that pair completely with Private mode and a shield icon, with a note, on the others.

The rule of thumb: an unreleased product, a client under NDA, a personal project you would rather nobody logged. Use an eroq engine with Private mode on.

References and editing

The moderated models read words and nothing else. On eroq they take no reference pictures and have no Edit mode, so each take invents its subject from your prompt. That is a strength when you are exploring and the wrong tool when you need this face, this bottle or this jacket.

Most of eroq's own engines read pictures:

• Eroq One takes up to ten references (faces, outfits, a product, a set) and edits by plain-language instruction: "swap the jacket for a red one", "make it night". 10 credits / image, +1 per extra picture, typically about 30 seconds.
• Krea2 and Eroq Uncensored take up to four references and edit a picture in place, for 10 credits / image each, typically about 15 seconds.
• Anima reads no pictures; it is the anime and illustration specialist, with a negative prompt and a second detail pass (see negative prompts explained).

The two families work well in sequence. Compose and explore on a moderated model, where range is the point, then carry the winning composition and your reference photos into Eroq One to hold identity. AI product photography in 2026 walks through that pattern for real products.

What each family will render

The moderated models each apply their own content rules before rendering. When one refuses a prompt, the render comes back as content_blocked and nothing is charged.

eroq's Eroq Uncensored, Eroq One and Anima are uncensored within the acceptable-use policy: intense fiction, dark themes and gritty realism render as written instead of being refused. Krea2 is the clean, everyday photoreal engine. The acceptable-use policy covers every model on eroq without exception, and a prompt involving minors is refused on every engine before any credit is charged.

Speed and price

The moderated models range from about 8 seconds (Nano Banana 2.1 and Flash Image) to about 25 seconds (Nano Banana Pro). eroq's engines run from about 15 seconds (Krea2, Eroq Uncensored) to about 30 (Eroq One) and about a minute for Anima, detail pass included. On price, eroq's own engines sit at the bottom of the lineup, while the October 2026 additions run from 31 credits / image up to 124 credits / image. The full lineup by job, price and speed is in the best AI image models of 2026.

Pick by situation

• Confidential concept work: Krea2 or Eroq One, Private mode on.
• A poster with a title: GPT Image 2.5 Sunburst; Private mode keeps the result unstored, though the prompt may be retained upstream.
• The same character across a series: Eroq One with reference pictures, or Krea2 for faster portraits.
• A dark thriller scene a moderated model turned down: Eroq Uncensored, within the acceptable-use policy.
• Anime or illustration: Anima.
• Maximum fidelity, accurate text, a complex brief: Nano Banana Pro.
• Texture-heavy landscapes and architecture: Seedream 5.0 Pro.
• Cheap, fast drafts: Flash Image at 10 credits / image, or Krea2 at 10 credits / image when the draft should stay private.

Developers get the same choice per request with the model field; the image API guide covers it. Open the image studio and check the badge before you render.

FAQ

Are GPT Image 2.5 and Nano Banana private on eroq?

They run on outside providers, which may retain your prompt but never use it for training. With Private mode on, eroq never stores the result. For fully private renders, use Krea2, Eroq Uncensored, Eroq One or Anima.

Do the providers train on my prompts?

No. The moderated models may retain a prompt but never use it for training, and the private engines (Krea2, Eroq Uncensored, Eroq One and Anima) keep nothing.

Which eroq image models are uncensored within the acceptable-use policy?

Eroq Uncensored, Eroq One and Anima, eroq's own engines, are uncensored within the acceptable-use policy. The other models are moderated, and Krea2 is the clean photoreal engine.

My prompt was blocked. Was I charged?

No. A prompt a moderated model refuses returns content_blocked and costs nothing, and so does a prompt that breaks the acceptable-use policy, on any engine. If the scene is within the policy, an eroq engine such as Eroq Uncensored can render it.

Can I use reference photos with Nano Banana Pro or Seedream?

Not on eroq: every moderated model is text-to-image only. For references, use Eroq One (up to ten pictures) or Krea2 and Eroq Uncensored (up to four).