Venice Image Generate

Choosing a model

curl "https://api.venice.ai/api/v1/models?type=image"

Inspect each model's model_spec:

• constraints.promptCharacterLimit — max prompt length (also applies to negative_prompt).
• constraints.aspectRatios[] + defaultAspectRatio — present on aspect-ratio-driven models; use aspect_ratio instead of width/height.
• constraints.resolutions[] + defaultResolution — present when the model accepts resolution.
• constraints.qualities[] + defaultQuality — present when the model accepts quality.
• constraints.steps.{default,max} — step bounds. Every model lists them, but only a few use steps (see the field table).
• constraints.widthHeightDivisor — pixel-sized models work best with width/height as multiples of this (8 or 16). The API does not validate it.
• supportsStyleReferences, constraints.maxStyleReferences, constraints.supportsStyleReferenceStrength — style-reference support.
• supportsWebSearch, supportsOptimizePromptThinking — whether those request flags do anything.
• privacy (private / anonymized) and uncensored — privacy tier and content posture.
• Pricing: pricing.generation.usd (flat per image), or pricing.resolutions[tier].usd for resolution-tiered models, plus pricing.quality[tier][level].usd for quality-tiered models.

Representative IDs (verify with GET /models?type=image — the list changes often):

bria-bg-remover also appears under type=image, but it is the background-removal model. Use it through /image/background-remove, not /image/generate.

Common patterns

Fixed-seed reproducibility

{"model": "z-image-turbo", "prompt": "...", "seed": 42}

On models that honor seed (e.g. z-image-turbo, seedream-v4, nano-banana-2), the same model + prompt + seed + settings should reproduce the same image, though third-party models don't guarantee bit-identical output. With variants > 1, only the first image uses seed; the rest are random, so run separate calls with different seeds if you need each one reproducible.

Aspect-ratio + resolution model (Nano Banana, Seedream V5 Pro)

{"model": "nano-banana-2", "prompt": "...", "aspect_ratio": "16:9", "resolution": "2K"}

{"model": "seedream-v5-pro", "prompt": "...", "aspect_ratio": "4:3", "resolution": "2K"}

Quality tier (GPT Image 2 / 2.5, Ideogram V4.5)

{"model": "gpt-image-2-5-flare", "prompt": "...", "aspect_ratio": "3:2", "resolution": "2K", "quality": "medium"}

Omitting quality uses defaultQuality (high for the GPT Image and Ideogram V4.5 models). The price depends on both resolution and quality.

Style preset + negative

Use a model that honors negative_prompt (see the field table); z-image-turbo silently drops it.

{
  "model": "venice-sd35",
  "prompt": "a red sports car in a parking lot",
  "negative_prompt": "blurry, people, clouds",
  "style_preset": "3D Model"
}

Style references (match the look of existing images)

{
  "model": "krea-v2-large",
  "prompt": "a lighthouse on a rocky coast at dusk",
  "style_references": [
    { "image": "https://example.com/ref-1.png", "strength": 0.8 },
    { "image": "data:image/png;base64,....", "strength": 0.4 }
  ]
}

Describe the subject in the prompt; the references carry the style. Today the supporting models are krea-v2-large / krea-v2-medium (up to 3 refs, strength honored) and luma-uni-1 / luma-uni-1-max (up to 3 refs, strength ignored). All four are anonymized models. Re-check supportsStyleReferences via GET /models?type=image. The Krea V2 models add a small per-request surcharge when references are used; it isn't itemized in the /models pricing.

Stream binary to disk (Node)

const res = await fetch('https://api.venice.ai/api/v1/image/generate', {
  method: 'POST',
  headers: { Authorization: `Bearer ${process.env.VENICE_API_KEY}`, 'Content-Type': 'application/json' },
  body: JSON.stringify({ model: 'z-image-turbo', prompt: '...', return_binary: true }),
})
if (!res.ok) throw new Error(await res.text())
if (res.headers.get('x-venice-is-content-violation') === 'true') throw new Error('Content violation')
const ext = (res.headers.get('content-type') ?? 'image/webp').split('/')[1]
const buf = Buffer.from(await res.arrayBuffer())
await fs.writeFile(`out.${ext}`, buf)

Errors

Content-policy violations are not an error on these endpoints. You get 200 with x-venice-is-content-violation: true and a blocked image, and no charge. See venice-errors for body shapes and retry strategy.

Gotchas

• Each model uses one sizing idiom: width/height, or aspect_ratio (+ resolution). Read constraints first. Sending width/height to an aspect-ratio model is silently ignored, except on qwen-image, qwen-image-3, and qwen-image-3-pro, where it is a 400.
• aspect_ratio isn't validated on /image/generate, so a value the model doesn't list usually falls back to its default without an error. A resolution or quality the model doesn't list is a 400.
• variants > 1 requires return_binary: false.
• Grok Imagine models return the provider's bytes unchanged when nothing is blurred, so the output format may not match format and EXIF isn't embedded. Trust Content-Type, or sniff the bytes.
• Always check x-venice-is-content-violation. A blocked image still arrives as a 200.
• style_references on a model without supportsStyleReferences: true is a 400, not a silent no-op.
• enhance_prompt bills $0.04 each time it produces a rewrite. Leave it off for cost- or latency-sensitive calls.
• For OpenAI-compat, response_format: "url" returns a data URL, not a hosted URL. Plan for that if you're saving to storage.