What is an NSFW image generation API?
An NSFW image generation API is a web endpoint you call from your own code to create images, hosted by a provider that allows adult output and runs models that don’t block it. You send a request with a prompt and some settings, and you get back an image, a URL to one, or a job ID that leads to one.
The difference from a mainstream image API is mostly in two places. The model itself has either been trained without the usual filtering or fine-tuned to produce adult content, and the provider doesn’t run a safety classifier that blanks out the result. Both matter. A permissive model behind a strict output filter still returns black squares, and a lenient provider can’t make a model draw what it never learned.
A typical request carries these fields, though names differ between providers:
- Model: the ID of the image model, such as an SDXL-based checkpoint or Qwen Image.
- Prompt and often a negative prompt listing what to avoid.
- Size or resolution tier, which usually changes the price.
- Seed, so you can reproduce an image or vary it slightly.
- Source image, only for image-to-image or editing.
If you already use an uncensored text model, some providers sell images on the same key. The comparison of uncensored LLM APIs covers the text side of those accounts.
Text-to-image vs image-to-image: which NSFW image API do you need?
Use text-to-image when your users describe what they want and expect a new picture. Use image-to-image when they start from a picture and want it changed: a new style, a different outfit, a cleaned-up render, or a variation of an earlier generation.
Text-to-image is the common case for generators, character art and companion apps that illustrate a chat. Image-to-image fits editing tools, upscale-and-restyle features, and pipelines where you generate one base image and then iterate on it. It also feeds video: most NSFW video APIs start from a still image, which is why the video API category lists image-to-video models.
| Provider | Text-to-image | Image-to-image | Payment | API format |
|---|---|---|---|---|
| Crebots | No | Yes | Crypto only | REST with API-KEY header |
| Uncensored.com | Yes | No | Card | Chat completions with X-Api-Key header |
| UncensoredLLM | Yes | Yes | Card or crypto | OpenAI and Anthropic |
| NanoGPT | Yes | No | Card or crypto | OpenAI and Anthropic |
Uncensored.com, UncensoredLLM and NanoGPT offer text-to-image. Crebots and UncensoredLLM offer image-to-image. Crebots is the specialist: it doesn’t generate from a prompt alone, but it transforms a source image at four resolution tiers and can carry the result on to video.
Which models do NSFW image APIs run?
Most hosted NSFW image models fall into three groups. Knowing the group tells you roughly what to expect before you spend money on a test.
- SDXL-based checkpoints. Stable Diffusion XL, released by Stability AI in 2023, generates at 1024 by 1024 pixels natively and has open weights. The community has produced thousands of fine-tuned versions of it, many aimed at photorealism or anime, and many trained on adult content. They are small and fast, which is why they’re usually the cheapest option. CyberRealistic XL Uncensored on UncensoredLLM is from this family.
- Newer large diffusion transformers. Chroma is an open model of about 8.9 billion parameters built from FLUX.1 [schnell] and released under the Apache 2.0 license. Qwen-Image is a 20-billion-parameter model from Alibaba’s Qwen team, notable for rendering text inside images and for editing. Models like these follow long prompts more closely, at a higher price per image.
- Edit models. Some models accept a source image and an instruction (“change the background to a beach”) rather than only a prompt. On UncensoredLLM, Qwen Image 2.1 Uncensored and Vision Image Uncensored also handle edits.
Anime and illustration models are a style choice rather than a different technology. Whatever the family, a provider’s model name tells you the base but not the fine-tune data, so test with your own prompts. The guide to how models get uncensored explains the same idea for text models.
How much does an NSFW image API cost per image?
NSFW image APIs in this directory charge between $0.0136 and $0.08 per image on published rates. Unlike text APIs, they bill per generation, so a short prompt and a long one cost the same. Price rises with model size and, where tiers exist, with resolution.
| Provider | Model or tier | Type | Per image | Per 1,000 images |
|---|---|---|---|---|
| UncensoredLLM | CyberRealistic XL Uncensored | Image generation | $0.0136 | $13.60 |
| Crebots | Low (~720p) | Image to Image | $0.02 | $20.00 |
| Crebots | Mid (~1080p) | Image to Image | $0.022 | $22.00 |
| Crebots | High (~1440p) | Image to Image | $0.024 | $24.00 |
| Crebots | Ultra (~4K) | Image to Image | $0.028 | $28.00 |
| UncensoredLLM | Anime Image Uncensored | Image generation | $0.03 | $30.00 |
| UncensoredLLM | Qwen Image 2.1 Uncensored (also edits) | Image generation | $0.0375 | $37.50 |
| UncensoredLLM | Vision Image Uncensored (also edits) | Image generation | $0.04 | $40.00 |
| UncensoredLLM | Chroma Uncensored | Image generation | $0.08 | $80.00 |
Uncensored.com and NanoGPT also offer image generation, but the directory doesn’t list per-image prices for them yet. Check the model list in your account before you compare. NanoGPT’s $12 a month Pro plan includes 100 images a day, which can beat per-image pricing for a small, steady workload.
The image-to-image tiers from Crebots are close together: Ultra (about 4K) costs $0.028, only 40% more than Low (about 720p), so resolution barely moves the bill there. Keep in mind those tiers need a source image, so they aren’t a direct substitute for text-to-image. On UncensoredLLM, the spread between models is much wider: the priciest image model costs about 6 times the cheapest.
Worked example: a monthly image budget
To budget an image app, count generations, not kept images. Users regenerate when a result is wrong, and every attempt is billed. Here is an app that delivers 10,000 images a month that users keep, with an average of 1.5 generations per kept image. That is 15,000 billed generations.
| Provider and model | Per image | Monthly cost |
|---|---|---|
| UncensoredLLMCyberRealistic XL Uncensored | $0.0136 | $204.00 |
| CrebotsLow (~720p) | $0.02 | $300.00 |
| CrebotsMid (~1080p) | $0.022 | $330.00 |
| CrebotsHigh (~1440p) | $0.024 | $360.00 |
| CrebotsUltra (~4K) | $0.028 | $420.00 |
| UncensoredLLMAnime Image Uncensored | $0.03 | $450.00 |
| UncensoredLLMQwen Image 2.1 Uncensored (also edits) | $0.0375 | $562.50 |
| UncensoredLLMVision Image Uncensored (also edits) | $0.04 | $600.00 |
| UncensoredLLMChroma Uncensored | $0.08 | $1,200.00 |
The choice of model moves the bill by $996.00 a month at this volume. Three things change these numbers in practice:
- Regeneration rate. If a model follows prompts badly, users retry more. A model that costs twice as much but halves retries costs the same in the end. Log attempts per kept image during a trial.
- Previews. Generating a cheap draft first and only rendering the final picture on a better model can cut costs, if your UI allows it.
- Free tiers for users. If you give away images, cap them. A free user who regenerates 50 times costs you the same as 50 paid images.
Prices come from the provider pages in this directory and can change. Confirm the current rate before you set your own prices.
How async image jobs work
Many image endpoints don’t return the picture in the same response. They return a job or task ID immediately, run the generation in a queue, and make the result available a few seconds to a minute later. Crebots documents this pattern with a task endpoint to poll and optional callbacks. Some OpenAI-compatible image endpoints return the image directly instead, so read the response format for the model you pick.
The flow for an async job looks like this:
- Submit. POST the prompt, model and settings. Store the returned job ID with the user’s request in your database.
- Wait. Either poll the status endpoint every few seconds, or register a callback URL that the provider calls when the job finishes. Callbacks are cheaper at scale. Polling is simpler to start with.
- Fetch. When the status is complete, download the image. Result URLs often expire, so copy the file to your own storage instead of linking to the provider’s URL.
- Handle failures. Set a timeout, treat “failed” as a final state, and decide whether a failed job is retried automatically or shown to the user.
A minimal polling loop in Python, with the endpoint paths left as placeholders because they differ between providers:
import os, time, requests
BASE = os.environ["IMAGE_API_BASE"]
HEADERS = {"API-KEY": os.environ["IMAGE_API_KEY"]}
job = requests.post(f"{BASE}/SUBMIT_PATH", headers=HEADERS,
json={"prompt": "...", "model": "..."}).json()
deadline = time.time() + 120
while time.time() < deadline:
status = requests.get(f"{BASE}/STATUS_PATH/{job['id']}", headers=HEADERS).json()
if status["status"] in ("completed", "failed"):
break
time.sleep(3)
The header name is an example too. Crebots uses an API-KEY header, Uncensored.com uses X-Api-Key, and OpenAI-compatible providers use a bearer token. Field names like id and status also vary, so map them from the provider’s documentation.
Consent, rights and moderation duties
An uncensored image model removes the model’s refusals. It doesn’t remove your legal duties as the app operator, and image apps carry more risk than text apps because they can depict real people.
- Minors. Sexual content involving minors is illegal everywhere, including drawn and AI-generated images in many countries. Block it in your prompts and outputs regardless of what the model allows. UncensoredLLM says it refuses any prompt involving a minor.
- Real people. Intimate images of real, identifiable people made without their consent are illegal in many places and against every provider’s terms. This matters most for image-to-image, where a user can upload someone else’s photo.
- Rights to source images. Only transform images you own or have permission to use. UncensoredLLM states that image edits must use images you have the rights to edit.
- Safety checkers. Some media models have a built-in safety checker setting. UncensoredLLM says it defaults to off on its media models that have one. That means filtering is your job, not the provider’s.
- Age gating and records. Adult platforms usually need age checks for users, and some laws require records about the content you host. Check the rules where you and your users are, and the provider’s terms.
A sensible setup runs your own checks on both the prompt and the output, keeps logs of generations, and gives users a way to report an image. None of the providers in this directory has been endpoint-tested by NSFW APIs, so read each one’s acceptable-use terms for your exact use case.
Checklist for choosing an NSFW image API
- Pick the request type. Text-to-image, image-to-image, editing, or a mix. That alone narrows the list.
- Shortlist two or three models. Include one cheap SDXL-based model and one larger model so you can see what the extra cost buys.
- Run the same 20 to 30 prompts on each. Score prompt following, anatomy errors and style. Note how many attempts each needs per usable image.
- Price the real workload. Multiply generations, not kept images, by the per-image rate at the resolution you need.
- Check the response format. Synchronous or async, how results are returned, and how long result URLs last.
- Check payment. Crebots takes crypto only, Uncensored.com takes cards only, and UncensoredLLM and NanoGPT take both.
- Plan moderation. Decide what you block, how you check uploads for image-to-image, and how users report problems, before launch.
Frequently asked questions
What is the cheapest NSFW image API?
On published rates in this directory, CyberRealistic XL Uncensored on UncensoredLLM is the cheapest at $0.0136 per image, or $13.60 per 1,000 images. Cheaper models may need more regenerations, so compare the cost per image your users actually keep.
Is there an NSFW AI image API for image-to-image?
Yes. Crebots and UncensoredLLM offer image-to-image in this directory. Crebots prices it by output resolution, from $0.02 to $0.028 per image.
Why does my NSFW image generator API return a job ID instead of an image?
The endpoint is asynchronous. The provider queues the generation and returns an ID straight away, and you poll a status endpoint or receive a callback to get the finished image.
Can an uncensored image generation API make images of real people?
Generating intimate or sexual images of real, identifiable people without their consent is illegal in many countries and banned by provider terms. Use only images and likenesses you have rights and consent to use.
Do NSFW image APIs charge per token?
No. Image APIs charge per generated image, and the price usually depends on the model and the output resolution. The length of the prompt doesn’t change the cost.