Docs

Running a model

Black Forest LabsText to image

Black Forest Labs FLUX.2 [klein] 9B Base

Undistilled foundation model for high-quality image generation and editing

POSTauray-serverless/bfl-flux-2-klein-9b-base
openapi.jsonllms.txt

One of Auray's catalogue models: the same key, the same credits, the same job lifecycle as every other address.

Its request shape is published below. Whether it runs right now is GET /v1/catalogue's word, with no key needed: its addresses say connected for an address that runs, and a submit to any other is refused before anything is charged — described answers 501 model_not_connected, paused 503 model_disabled.

Category text-to-image · Scope image:write (or external:write on an older key) · Price Credits per megapixel (counted as width*height*steps@28): 0.9689 for acceleration=none; acceleration=low; acceleration=medium; acceleration=high. From 1.0322 to 7.4899 megapixels a request. Quoted when you submit, at the most the request can cost; the charge settles to what the job cost and never above the quote. The submit's answer carries the quote as credits_quoted, and credits_charged says what has been taken.

#Input

FieldTypeRequiredDefaultDescription
inputsobjectno—
inputs.referenceImagesstring[]no—List of reference images (an http(s) URL; upload a local file first and send its read URL). Documented range: 1 to 4.
positivePromptstringyes—Text prompt describing elements to include in the generated output. Documented range: 1 to 10,000 characters.
negativePromptstringno—Prompt to guide what to exclude from generation. Ignored when guidance is disabled (CFGScale ≤ 1). Documented range: 1 to 10,000 characters.
widthintegeryes1024Width of the generated media in pixels. Documented range: 128 to 2,048.
heightintegeryes1024Height of the generated media in pixels. Documented range: 128 to 2,048.
seedintegerno—Random seed for reproducible generation. When not provided, a random seed is generated in the unsigned 32-bit range. Documented range: 0 to 9,223,372,036,854,776,000.
stepsintegerno28Total number of denoising steps. Higher values generally produce more detailed results but take longer. Documented range: 1 to 50.
schedulerstringno—Scheduler to use for the diffusion process. One of "DDIM", "DDIMScheduler", "DDPMScheduler", "DEISMultistepScheduler", "Default", "DPM++", "DPM++ 2M", "DPM++ 2M Beta", "DPM++ 2M Exponential", "DPM++ 2M Karras", "DPM++ 2M SDE", "DPM++ 2M SDE Beta", "DPM++ 2M SDE Exponential", "DPM++ 2M SDE Karras", "DPM++ 2M SDE Uniform", "DPM++ 2M Uniform", "DPM++ 3M", "DPM++ 3M Beta", "DPM++ 3M Exponential", "DPM++ 3M Karras", "DPM++ 3M SDE Uniform", "DPM++ 3M Uniform", "DPM++ Beta", "DPM++ Exponential", "DPM++ Karras", "DPM++ SDE", "DPM++ SDE Beta", "DPM++ SDE Exponential", "DPM++ SDE Karras", "DPM++ Uniform", "DPM++ Uniform Beta", "DPM++ Uniform Exponential", "DPM++ Uniform Karras", "DPMSolverMultistepInverse", "DPMSolverMultistepScheduler", "DPMSolverSinglestepScheduler", "EDMDPMSolverMultistepScheduler", "EDMEulerScheduler", "Euler", "Euler a", "Euler Beta", "Euler Exponential", "Euler Karras", "EulerAncestralDiscreteScheduler", "EulerDiscreteScheduler", "FlowMatchEulerDiscreteScheduler", "Heun", "HeunDiscreteScheduler", "Heun Karras", "IPNDMScheduler", "IPNDM Uniform", "IPNDM Uniform Beta", "IPNDM Uniform Exponential", "IPNDM Uniform Karras", "KDPM2AncestralDiscreteScheduler", "KDPM2DiscreteScheduler", "LCM", "LCMScheduler", "LMS", "LMSDiscreteScheduler", "LMS Karras", "PNDMScheduler", "TCDScheduler", "UniPC", "UniPC 2M", "UniPC 2M Karras", "UniPC 2M Uniform", "UniPC 3M", "UniPC 3M Karras", "UniPC 3M Uniform", "UniPC Karras", "UniPC Uniform", "UniPC Uniform Beta", "UniPC Uniform Exponential", "UniPC Uniform Karras".
CFGScalenumberno3.5Guidance scale representing how closely the output will resemble the prompt. Higher values produce results more aligned with the prompt. Documented range: 1 to 20.
accelerationstringno"high"Optimization level. One of "none", "low", "medium", "high".
acceleratorOptionsobjectno—
acceleratorOptions.cacheEndStepintegerno—Absolute step number to end caching. Must be greater than cacheStartStep and less than or equal to steps. Documented range: 1 or more.
acceleratorOptions.cacheEndStepPercentageintegerno—Percentage of steps to end caching. Alternative to cacheEndStep. Must be greater than cacheStartStepPercentage. Documented range: 1 to 100.
acceleratorOptions.cacheMaxConsecutiveStepsintegerno3Limits the maximum number of consecutive steps that can use cached computations before forcing a fresh computation. Documented range: 1 to 5.
acceleratorOptions.cacheStartStepintegerno—Absolute step number to start caching. Must be less than cacheEndStep. Documented range: 0 or more.
acceleratorOptions.cacheStartStepPercentageintegerno—Percentage of steps to start caching. Alternative to cacheStartStep. Must be less than cacheEndStepPercentage. Documented range: 0 to 99.
acceleratorOptions.teaCachebooleanno—TeaCache acceleration for transformer-based models. Estimates step differences to skip redundant computations.
acceleratorOptions.teaCacheDistancenumberno0.5Controls the aggressiveness of the TeaCache feature. Lower values prioritize quality, higher values prioritize speed. Documented range: 0 to 1.
acceleratorOptions.dbCachebooleanno—DB Cache (CacheDiT) acceleration. Caches and reuses intermediate transformer block outputs to skip redundant computations.
acceleratorOptions.dbCacheThresholdnumberno0.25Controls the sensitivity threshold for DB Cache. Lower values reuse cached blocks more aggressively, higher values prioritize quality. Documented range: 0 to 1.
acceleratorOptions.dbCacheSkipIntervalintegerno5Controls how many steps to skip between cache refreshes. Higher values skip more steps for faster generation at the cost of quality. Documented range: 1 or more.
loraobject[]no—
lora[].modelstringno—LoRA model identifier.
lora[].weightnumberno1Strength of the LoRA influence. A value of 0 means no influence. Higher values increase the influence, and negative values can be used to steer away from the LoRA's style. Documented range: -4 to 4.
outputTypestringno"URL"Image output type. One of "URL".
outputFormatstringno"JPG"Specifies the file format of the generated output. The available values depend on the task type and the specific model's capabilities. - `JPG`: Best for photorealistic images with smaller file sizes (no transparency). - `PNG`: Lossless compression, supports high quality and transparency (alpha channel). - `WEBP`: Modern format providing superior compression and transparency support. - `MP4`: Widely supported video container (H.264), recommended for general use. - `WEBM`: Optimized for web delivery. - `MOV`: QuickTime format, common in professional workflows (Apple ecosystem). - `GIF`: Animated image format (no audio), suitable for short loops or previews. - `MP3`: Compressed audio, smaller file size. - `WAV`: Uncompressed, high-quality audio. - `FLAC`: Lossless compression. - `OGG`: Open-source compressed audio format (Vorbis codec). - `SVG`: Scalable Vector Graphics. - `TIFF`: High-quality output supporting layers. - `GLB`: Binary glTF format for 3D models, including geometry, textures, and PBR materials. > [!NOTE] > **Transparency**: If you are using features like background removal or LayerDiffuse that require transparency, you must select a format that supports an alpha channel (e.g., `PNG`, `WEBP`, `TIFF`). `JPG` does not support transparency. One of "JPG", "PNG", "WEBP".
outputQualityintegerno95Compression quality of the output. Higher values preserve quality but increase file size. Documented range: 20 to 99.
safetyobjectno—
safety.checkContentbooleanno—Enable or disable content safety checking. Increases total generation time. Always on: a false is sent as true.
advancedFeaturesobjectno—
advancedFeatures.watermarkobjectno—
advancedFeatures.watermark.textstringno—Watermark text. Documented range: 2 to 32 characters.
advancedFeatures.watermark.imagestringno—Watermark image (an http(s) URL; upload a local file first and send its read URL). A URL.
advancedFeatures.watermark.displayPositionstringno—Watermark position. One of "top-left", "top-center", "top-right", "center-left", "center-center", "center-right", "bottom-left", "bottom-center", "bottom-right", "tiled".
advancedFeatures.watermark.opacitynumberno—Watermark opacity. Documented range: 0.1 to 1.
advancedFeatures.watermark.fontColorstringno—Text color in hex format.
advancedFeatures.watermark.bgColorstringno—Background color in hex format.
idempotency_keystringno—Your own name for this request. Send the same one twice and the second call is answered with the first request rather than charged again. Idempotency-Key: as a header is accepted as an alias — that is the spelling a retrying proxy replays, and the one to use when running a model on auray.run, where a request that outlives its connection is answered 504 and recovered by replay. Omit it and the platform mints one. 1 to 64 characters.

#Rules

Across fields, as the model's documentation states them. Each is judged before anything is charged — warned with a Deprecation header until enforcement starts — and listed as JSON Schema under x-auray-rules in the openapi.json.

  • Send height whenever width is sent. Send width whenever height is sent.
  • Do not send acceleration together with acceleratorOptions.
  • Do not send acceleratorOptions.cacheStartStep and acceleratorOptions.cacheStartStepPercentage together. Do not send acceleratorOptions.cacheEndStep and acceleratorOptions.cacheEndStepPercentage together. Do not send a request where acceleratorOptions.teaCache is true and acceleratorOptions.dbCache is true.
  • lora must hold at least 1 item. In every item of lora: lora[].model is required.
  • When advancedFeatures.watermark is sent: send exactly one of advancedFeatures.watermark.text or advancedFeatures.watermark.image.

#Example

curl -X POST https://queue.auray.run/auray-serverless/bfl-flux-2-klein-9b-base \
  -H "Authorization: Bearer $AURAY_API_KEY" \
  -H "content-type: application/json" \
  -d '{
  "positivePrompt": "a brass orrery on a walnut desk, morning light from a window to the left",
  "width": 1024,
  "height": 1024
}'

The same body on https://auray.run runs synchronously and answers the result — unless it usually takes longer than a run may hold its connection, which is refused 400 model_not_synchronous before anything is charged. Options — ?webhook=, ?wait=, priority and timeout headers — are on Submit a request.

#Machine contracts

  • openapi.json — this page as an OpenAPI 3.1 document, public.
  • llms.txt — the same, written for an agent.

Start typing to search every page of the documentation.

Up arrow Down arrow to moveEnter to openEscape to close