MODEL IDrunware:400@6

live

FLUX.2 [klein] 9B KV

by Black Forest LabsMarch 20, 2026

FLUX.2 [klein] 9B KV is a KV-cache optimized variant of the Klein 9B model that caches reference image key-value pairs after the first denoising step, skipping redundant computation on subsequent steps. This delivers up to 2.5x faster inference for multi-reference editing tasks while retaining all capabilities of the standard Klein 9B, including sub-second text-to-image and advanced editing in 4 steps.

API Options

Platform-level options for task execution and delivery.

taskType stringrequiredvalue: imageInference: Identifier for the type of task being performed

taskUUID stringrequiredUUID v4: UUID v4 identifier for tracking tasks and matching async responses. Must be unique per task.

outputType stringdefault: URL: Image output type.
Allowed values3 values

outputFormat stringdefault: 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.

**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.

Allowed values3 values

outputQuality integermin: 20max: 99default: 95: Compression quality of the output. Higher values preserve quality but increase file size.

webhookURL stringURI

Specifies a webhook URL where JSON responses will be sent via HTTP POST when generation tasks complete. For batch requests with multiple results, each completed item triggers a separate webhook call as it becomes available.

Learn more1 resource

Webhooks
PLATFORM

deliveryMethod stringdefault: sync

Determines how the API delivers task results.

Allowed values2 values

: Returns complete results directly in the API response.
: Returns an immediate acknowledgment with the task UUID. Poll for results using getResponse.

Learn more1 resource

Task Polling
PLATFORM

uploadEndpoint stringURI

Specifies a URL where the generated content will be automatically uploaded using the HTTP PUT method. The raw binary data of the media file is sent directly as the request body. For secure uploads to cloud storage, use presigned URLs that include temporary authentication credentials.

Common use cases:

Cloud storage: Upload directly to S3 buckets, Google Cloud Storage, or Azure Blob Storage using presigned URLs.
CDN integration: Upload to content delivery networks for immediate distribution.

// S3 presigned URL for secure upload
https://your-bucket.s3.amazonaws.com/generated/content.mp4?X-Amz-Signature=abc123&X-Amz-Expires=3600

// Google Cloud Storage presigned URL
https://storage.googleapis.com/your-bucket/content.jpg?X-Goog-Signature=xyz789

// Custom storage endpoint
https://storage.example.com/uploads/generated-image.jpg

The content data will be sent as the request body to the specified URL when generation is complete.

safety object

Content safety checking configuration for image generation.

Properties1 property

safety » checkContent checkContent booleandefault: false: Enable or disable content safety checking.

ttl integermin: 60: Time-to-live (TTL) in seconds for generated content. Only applies when outputType is URL.

includeCost booleandefault: false: Include task cost in the response.

Inputs

Input resources for the task (images, audio, etc). These must be nested inside the inputs object.

inputs » referenceImages referenceImages array of stringsmin items: 1max items: 4: List of reference images (UUID, URL, Data URI, or Base64).

Core Parameters

Primary parameters that define the task output.

numberResults integermin: 1max: 20default: 1: Number of results to generate. Each result uses a different seed, producing variations of the same parameters.

model stringrequiredvalue: runware:400@6: Identifier of the model to use for generation.

positivePrompt stringrequiredmin: 1max: 10000: Text prompt describing elements to include in the generated output.

negativePrompt stringmin: 2max: 3000: Prompt to guide what to exclude from generation. Ignored when guidance is disabled (CFGScale ≤ 1).

width integerrequiredmin: 128max: 2048step: 16: Width of the generated media in pixels.

height integerrequiredmin: 128max: 2048step: 16: Height of the generated media in pixels.

seed integermin: 0max: 9223372036854776000: Random seed for reproducible generation. When not provided, a random seed is generated in the unsigned 32-bit range.

steps integermin: 1max: 50default: 4: Total number of denoising steps. Higher values generally produce more detailed results but take longer.

scheduler string: Scheduler to use for the diffusion process.
Allowed values75 values

CFGScale floatmin: 1max: 20step: 0.01default: 3.5: Guidance scale representing how closely the output will resemble the prompt. Higher values produce results more aligned with the prompt.

acceleration stringdefault: high: Optimization level.
Allowed values4 values

acceleratorOptions object

Advanced caching mechanisms to speed up generation.

Properties12 properties

acceleratorOptions » cacheEndStep cacheEndStep integermin: 1: Absolute step number to end caching. Must be greater than cacheStartStep and less than or equal to steps.

acceleratorOptions » cacheEndStepPercentage cacheEndStepPercentage integermin: 1max: 100: Percentage of steps to end caching. Alternative to cacheEndStep. Must be greater than cacheStartStepPercentage.

acceleratorOptions » cacheMaxConsecutiveSteps cacheMaxConsecutiveSteps integermin: 1max: 5default: 3: Limits the maximum number of consecutive steps that can use cached computations before forcing a fresh computation.

acceleratorOptions » cacheStartStep cacheStartStep integermin: 0: Absolute step number to start caching. Must be less than cacheEndStep.

acceleratorOptions » cacheStartStepPercentage cacheStartStepPercentage integermin: 0max: 99: Percentage of steps to start caching. Alternative to cacheStartStep. Must be less than cacheEndStepPercentage.

acceleratorOptions » fbCache fbCache booleandefault: false: First Block Cache (FBCache) acceleration. Reuses feature block computations across steps.

acceleratorOptions » fbCacheThreshold fbCacheThreshold floatmin: 0max: 1step: 0.01default: 0.25: Controls the sensitivity threshold for determining when to reuse cached computations. Lower values reuse more aggressively.

acceleratorOptions » teaCache teaCache booleandefault: false: TeaCache acceleration for transformer-based models. Estimates step differences to skip redundant computations.

acceleratorOptions » teaCacheDistance teaCacheDistance floatmin: 0max: 1step: 0.01default: 0.5: Controls the aggressiveness of the TeaCache feature. Lower values prioritize quality, higher values prioritize speed.

acceleratorOptions » dbCache dbCache booleandefault: false: DB Cache (CacheDiT) acceleration. Caches and reuses intermediate transformer block outputs to skip redundant computations.

acceleratorOptions » dbCacheThreshold dbCacheThreshold floatmin: 0max: 1step: 0.01default: 0.25: Controls the sensitivity threshold for DB Cache. Lower values reuse cached blocks more aggressively, higher values prioritize quality.

acceleratorOptions » dbCacheSkipInterval dbCacheSkipInterval integermin: 1default: 5: Controls how many steps to skip between cache refreshes. Higher values skip more steps for faster generation at the cost of quality.

lora array of objectsmin items: 1

Configuration for Low-Rank Adaptation models.

Properties3 properties

lora » model model stringrequired: LoRA model identifier.

lora » weight weight floatmin: -4max: 4step: 0.01default: 1: Strength 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.

lora » transformer transformer stringdefault: both

Transformer stages to apply LoRA. Some video models use separate high-noise and low-noise processing stages, and LoRAs can be selectively applied to optimize their effectiveness.

Allowed values3 values

: Apply LoRA only to the high-noise processing stage (coarse structure and early generation steps).
: Apply LoRA only to the low-noise processing stage (fine details and later generation steps).
: Apply LoRA to both stages for full coverage.

Features

Standalone addons and post-processing features.

ultralytics object

Configuration object for Ultralytics face enhancement during generation. This feature uses face detection and inpainting to improve facial details in the same generation step, without requiring post-processing.

Face enhancement is available for Stable Diffusion 1.X, SDXL, and FLUX models. The system automatically detects faces and applies targeted refinement to improve quality while maintaining consistency with the overall generation.

Properties9 properties

ultralytics » CFGScale CFGScale floatmin: 0max: 50step: 0.1default: 8: Face refinement guidance scale.

ultralytics » confidence confidence floatmin: 0max: 1step: 0.01default: 0.9: Confidence threshold for detection.

ultralytics » inpaintSize inpaintSize integermin: 128max: 2048default: 1024: Image size (in pixels) to use for each inpainting region. YOLO detects faces, crops the region, and scales it to this size before running diffusion. Set so most faces land in the 2–4× range of their original pixel size. Going beyond 8× may degrade identity resemblance.

ultralytics » maskBlur maskBlur integermin: 0max: 100default: 5: Mask feathering amount. Higher values create softer transitions between the enhanced face region and surrounding areas.

ultralytics » maskPadding maskPadding integermin: 0max: 20default: 5: Padding around detected face in pixels. Expands the refinement area to include surrounding context like hair and neck.

ultralytics » negativePrompt negativePrompt string: Negative prompt for detection.

ultralytics » positivePrompt positivePrompt string: Positive prompt for detection.

ultralytics » steps steps integermin: 1max: 100default: 20: Number of face refinement steps.

ultralytics » strength strength floatmax: 1step: 0.01default: 0.3: Refinement strength. Lower values preserve more of the original, higher values allow more aggressive reconstruction.