NovelAI Image Generation

rout.my exposes the full NovelAI Diffusion image catalog through the standard OpenAI-compatible images endpoint. All 144 working models are available under the novelai/ prefix — no separate account, no NovelAI subscription.

Quick start

curl https://api.rout.my/v1/images/generations \
  -H "Authorization: Bearer sk_..." \
  -H "Content-Type: application/json" \
  -d '{
    "model": "novelai/nai-diffusion-4-5-full-1024x1024-s28",
    "prompt": "a serene mountain lake at dawn, mist over the water",
    "n": 1
  }'

The response is the standard OpenAI images shape: data[0].b64_json contains the PNG.

Model naming

The full generation parameters are embedded in the model ID:

novelai/nai-diffusion-<version>-<WxH>-s<steps>

Part	Meaning	Examples
version	model family	4-full, 4-5-curated, 5-full, 4-curated-preview
WxH	output resolution	1024x1024, 1216x832, 1536x1536
s<steps>	sampling steps	s20 (fast), s24 (balanced), s28 (quality)

Available resolutions: 832x1216, 1216x832, 1024x1024, 1024x1536, 1536x1024, 1344x768, 768x1344, 1536x1536.

Six families × eight sizes × three step counts = 144 models. Browse the full list in the model catalog — families with all variants healthy are shown as a single row; click it to expand.

Choosing a model

• nai-diffusion-4-5-full — the latest full model, best overall quality.
• nai-diffusion-4-5-curated — the curated variant, more aesthetic-safe defaults.
• nai-diffusion-5-full / nai-diffusion-5-curated — newest generation.
• nai-diffusion-4-curated-preview — older preview line, cheapest tier.
• Steps: s20 for drafts and iterating on prompts, s28 for final renders.

Request parameters

{
  "model": "novelai/nai-diffusion-4-5-full-1024x1024-s28",
  "prompt": "a serene mountain lake at dawn",
  "negative_prompt": "blurry, lowres, watermark",
  "n": 1
}

Parameter	Type	Notes
model	string	Required. Full model ID with the novelai/ prefix.
prompt	string	Required. What to generate.
negative_prompt	string	Optional. Appended to the prompt as a negative directive — the upstream accepts it verbatim.
n	integer	Images per request.

WARNING

Resolution and step count are part of the model ID — there is no separate size or steps parameter. Pick the exact variant you need from the catalog.

Pricing

Each model bills per generated image. The multiplier scales with the generation cost of the variant: the cheapest tier (1024x1024, s20, curated preview) is 5.0×, and larger sizes plus more steps multiply it — a 1536x1536 s28 full model reaches 202.0× per image. Check the exact multiplier next to every model on the dashboard catalog.

All NovelAI models require a paid plan.

Image references (img2img)

NovelAI accepts a single reference image per request through the chat-completions endpoint. Send the image as a standard OpenAI vision image_url content block (a data: URI or a hosted URL) in the user message together with your prompt:

{
  "model": "novelai/nai-diffusion-4-curated-preview-1024x1024-s20",
  "messages": [
    {
      "role": "user",
      "content": [
        {"type": "text", "text": "redraw this character in watercolor"},
        {"type": "image_url", "image_url": {"url": "data:image/png;base64,..."}}
      ]
    }
  ]
}

The upstream enforces exactly one image per request — sending two or more returns HTTP 400. The error is automatically translated into a readable message, so your client will see "NovelAI image reference supports exactly one input image per request" instead of the upstream's raw text.

In SillyTavern this maps to attaching one image to the message (via the paperclip or [img] tag) while using a chat-completion source — the same flow as any vision model.

Using with a frontend

Any client that speaks the OpenAI images API works. Point it at https://api.rout.my/v1, set your key, and pick a novelai/... model ID.

For SillyTavern, set up a custom image generation backend with the images endpoint above — see the SillyTavern guide for the general Chat Completion setup, then use /v1/images/generations for image calls.

Limits and behavior

• One request = one image per n; batching is supported.
• One image reference per img2img request (see the section above).
• The upstream rejects streaming requests, so generation is synchronous — expect a few seconds per image depending on steps and size.
• Old nai-diffusion-3 and nai-diffusion-furry-3 families are intentionally not offered — their upstream channel is unreliable.