Unfiltered image generation API

What passes through untouched

The design rule is that the platform does not have opinions about your prompt. There is no language model in front of the sampler rewriting your text into something it considers better, and no scoring pass between the finished pixels and the URL you get back.

• Your prompt — the exact string you send is the string the sampler conditions on. It is not expanded, paraphrased, or "improved".
• Your negative prompt — nothing is appended. No SFW rating tags, no per-model negative suffix, no house style.
• Your parameters — steps, guidance, seed, scheduler and size are passed as given, so a fixed seed reproduces a fixed image.
• The output — generated images and sampled video frames are not classified before the URL is returned.

The two strings that are prepended, and why

Full disclosure, because "verbatim" should mean something exact. Two tokens can appear ahead of your prompt, and both are mechanical rather than editorial.

Neither is a content judgement, and neither is a secret — the model page for each checkpoint states its prefix, and a LoRA trigger is documented alongside the LoRA. If you send no LoRA, only the checkpoint prefix applies.

• The checkpoint’s native quality tags — the short token string that particular checkpoint was trained to expect. Omitting it does not censor anything, it just produces worse images on that model.
• A catalog LoRA’s trigger word, and only when you select that LoRA — the trigger is the mechanism by which a LoRA is invoked at all. No LoRA selected, no trigger added.

The one gate that is enforced

The positive prompt is matched against a local minor-safety term list before the job is queued, and a prompt combining a minor with sexual content is rejected. This runs in-process against a deterministic list — no third-party classifier, no vendor call, no copy of your prompt leaving the platform. The negative prompt is deliberately not scanned for these terms, because an age word in a negative is a safety measure rather than a request.

It is scoped to the request and nothing else. The call returns 422, nothing is charged, and the account is untouched: no strike, no pause, no key revocation. This is an illegal-content floor, not a behaviour score.

{
  "error": {
    "type": "content_blocked",
    "code": "content_blocked_prompt",
    "message": "Request blocked by content policy (prompt). See https://fastgencloud.com/legal/aup",
    "request_id": "req_…"
  }
}

Why prompt rewriting is a product problem

Rewriting is usually described as a quality feature, and for a single interactive user it sometimes is. Inside a product it breaks things that are difficult to work around.

• Reproducibility disappears. If the platform rewrites your prompt non-deterministically, a fixed seed no longer fixes the image, and "regenerate" stops meaning anything.
• You cannot A/B test prompt wording, because the thing you are testing is not the thing being run.
• Debugging gets indirect. A bad result might be your prompt or might be the rewrite, and you cannot see which.
• Your users’ intent gets edited by a third party that has never seen your product or your terms of service.
• Silent style drift. Injected rating tags and house negatives push every output toward one aesthetic, which is the opposite of what a white-label product wants.

Check it yourself

The job record echoes back the prompt that was submitted, so you can diff what you sent against what was stored. Submit a prompt with a deliberate typo and a fixed seed, then fetch the job: the typo is still there, and generating twice with the same seed returns the same image.

curl https://api.fastgencloud.com/v1/images/generations \
  -H "Authorization: Bearer $FASTGEN_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
      "model": "general-image",
      "prompt": "a lone figure on a rain-slicked street at night, neon reflections, 35mm",
      "negative_prompt": "blurry, extra fingers",
      "seed": 1234567,
      "size": "1024x1024"
    }'

# then read it back — prompt and negative_prompt are stored as sent
curl https://api.fastgencloud.com/v1/jobs/{id} -H "Authorization: Bearer $FASTGEN_API_KEY"

Your product, your policy

The split below is the whole content model. The platform enforces the legal floor and nothing above it; everything else is a decision you make for your own users, and the Acceptable Use Policy is what you agree to when you make it.

If you expose generation to end users, you are the publisher. That means an age gate where your jurisdiction requires one, consent for real people’s likenesses, and your own moderation on whatever you choose to display or store.

Sexual content involving minors

Blocked at the prompt. Always, not configurable.

Adult content between adults

Passes. Your age gate, your terms, your call.

Violence, horror, gore

Passes. Not scored, not blocked.

Political and satirical content

Passes. No topic list.

Real people’s likenesses

Passes technically; restricted by the AUP and enforced on report.

Output retention

Deleted after 7 days. Copy anything you need to keep.