Black Forest Labs

FLUX 3 Image Klein

flux-3-image-klein

Try flux-3-image-klein

Input

Showcases

View all showcases

Output

image

Fill in the parameters and run — results show up here.

Pricing

flux-3-image-klein per image

Specifications	Unit Price
768	$0.0092
1K	$0.0103
1.5K	$0.0126
2K	$0.016
4K	$0.0412

Positioning Overview

This is a model designed for image generation and editing scenarios. It supports generating images from prompts alone, and can also modify or composite a scene according to instructions when given reference images. Input can include up to 10 reference images, and tags together with bounding boxes can be used to describe layout and local regions. The output is a static image, with optional 768 to 4K pixel tiers and a variety of landscape and portrait aspect ratios. It is suitable for developers and design teams that need to produce visual assets in batches, continue the style of reference images, or control layout.

Key Highlights

Text-to-image and reference-image editing combined — The same model accepts both pure prompt-based generation and reference images plus instructions for modification and compositing.
Combine up to ten reference images — Multiple reference images can be passed in a single request, for multi-image compositing or assigning different subject and style roles to different images.
Output tiers up to 4K — Resolution can be selected from 768, 1K, 1.5K, 2K, and 4K, with width and height changing according to the selected aspect ratio.
Bounding-box layout control — In the prompt, tags together with bounding-box JSON can specify element positions and edit local regions.

Capability Description

Image generation from prompts alone: Submit only a prompt to produce an image; prompts can be up to 5000 characters. When no reference image is passed, it is text-to-image mode, suitable for building a scene from scratch.

Instruction-based editing with reference images: Pass reference images and describe in the prompt what should be modified and preserved to make targeted adjustments to the scene; up to 10 reference images are supported.

Multiple reference images and layout control: Multiple reference images can serve as subject, style, or composition references, and tags plus bounding-box JSON can specify element placement, making it easier to organize complex scenes.

Aspect ratio and format selection: Aspect ratios can be selected from 21:9 to 9:21 across multiple landscape and portrait compositions, or automatic mode; output formats are jpeg or png.

Adjustable steps and random seed: Diffusion steps are adjustable from 1 to 16; fewer steps generate faster but may reduce detail. Specifying a seed from 0 to 4294967295 helps reproduce results under the same parameters as closely as possible.

Use Cases

E-commerce and marketing asset retouching — Operations staff upload product reference images with modification instructions to obtain adjusted main images or variants for product detail pages and campaign asset slots.
Multi-reference style compositing — Designers pass in a subject image and style reference images to composite visual drafts with a unified tone for brand promotional images and social media visuals.
Poster generation with layout control — Marketing teams use prompts together with bounding boxes to describe title and element positions, producing horizontal or vertical ad images for delivery and print previews.
High-resolution final image output — Content teams select 2K or 4K tiers to generate final images for display pages and material output that require higher pixel dimensions.
Batch image prototype validation — Developers use fixed seeds and lower step counts to quickly generate multiple sets of sketches for creative direction screening and early concept comparison.

Limitations and Selection Advice

Input prompts must be 1 to 5000 characters, and reference images must be accessible http or https image links, up to 10. The resolution, aspect ratio, and output format of each request are each selected as one option. The maximum diffusion step count is 16; lower step counts may reduce detail. The resolution tier describes the output pixel tier, and the specific width and height vary with the aspect ratio. The seed is only used to reproduce results under the same parameters as closely as possible and is not guaranteed to be completely identical; the content safety tolerance ranges from 0 to 4, with 0 being the strictest.

What's new

View changelog →

New model FLUX 3 Image Klein Now Available: Text-to-Image and Reference Editing Nov 11, 2026

Frequently asked questions

View all FAQs →

Why do short prompts produce generic images?

Short prompts give the model too little information, so it can only fill in the gaps based on common interpretations, which makes the results feel generic. Writing the prompt as a complete description is far more reliable: ① subject and appearance details ② scene and environment ③ lighting and color tone ④ composition and viewpoint ⑤ style and image-quality requirements. The prompt limit is 5000 characters, which is enough to spell out each item clearly.
How do I switch between text-to-image and image editing?

The same model supports two usage modes; the only difference is whether reference images are passed. Omitting image_urls means text-to-image, generating solely from the prompt; passing image_urls switches to instruction-based editing, where the prompt must clearly state which parts of the reference image need to be modified and which need to be preserved. When editing, it is recommended to list each part to be changed and the desired effect one by one, rather than writing just a single sentence about the overall style.
How many reference images can be used at most?

image_urls supports a maximum of 10 images and a minimum of 1; once reference images are passed in, it enters edit mode or multi-reference composition mode. For each image, you need to explain its role in the prompt, for example, the first image is the character, the second is the product, and the third is the background style. This upper limit is suitable for composition scenarios that require cross-image consistency of characters or products.
How do I control layout composition?

Use the labels in the prompt together with the bbox JSON to specify element positions. In the prompt, write placeholders in the form <label> and attach the corresponding bounding-box JSON to state which region each element goes in, then write the specific description into the content corresponding to each label. This is suitable for images such as posters and banners that require fixed layout positions.

Create a generation task

POST https://api.hiapi.ai/v1/tasks

Headers

AuthorizationstringRequired

Account API key. Required on every request. Example: Bearer YOUR_HIAPI_API_KEY. Get your API key

Content-TypestringOptional

The request body is always parsed as JSON; when sent, only application/json (optionally with charset) is accepted — any other value returns 415. Example: application/json.

Idempotency-KeystringOptional

Idempotency key, up to 255 bytes. Retries with the same key from the same account create the task only once; replays return the original taskId.

Idempotency key

On network timeouts or client retries, the same key creates the task only once — no double charges (scope: the API key's account). Use a business-unique identifier as the key (e.g. "order-42"); replays return the original taskId, and keys expire 24 hours after the first request. See the Unified Async API overview for the full scenario matrix.

Body

modelstringRequired

Current fixed value: flux-3-image-klein

routestringOptional

Current fixed value: default

inputobjectRequired

Business parameters for flux-3-image-klein and route default — see the parameter table below.

seedintegerOptionalmin 0max 4294967295

Random seed used to reproduce the same result as closely as possible under identical parameters; takes an integer, randomly generated when omitted.

stepsintegerOptionalDefault: 16min 1max 16

Number of diffusion steps, ranging from 1 to 16, default 16; lower values generate faster and may reduce detail, without affecting price.

promptstringRequiredminLength 1maxLength 5000

Describe the image you want to generate, or explain what needs to be modified and preserved in the reference images; can be combined with <tags> and bbox JSON to specify layout and local editing regions.

image_urlsarrayOptionalminItems 1maxItems 10

Array of reference images, each element being an image URL; the subject, style, or compositional role of each image must be explained in the prompt; omitted means text-to-image.

Array [ string ]

resolutionstringOptionalDefault: "1K"

Output pixel tier; the specific width and height vary with the aspect ratio; default 1K.

Options:7681K1.5K2K4K

aspect_ratiostringOptionalDefault: "auto"

Output aspect ratio, can be landscape, portrait, or square; default auto, where automatic mode determines the aspect ratio based on the prompt and reference images.

Options:21:92:116:93:27:54:35:41:1

Show all (8)Show less

4:53:45:72:39:161:29:21auto

output_formatstringOptionalDefault: "jpeg"

Output image format, default jpeg.

Options:jpegpng

safety_toleranceintegerOptionalDefault: 2min 0max 4

Content safety tolerance, takes an integer, 0 being the strictest; default 2.

callbackobjectOptional

Terminal-state notification config. When omitted, no callback is delivered — poll the task yourself. See the callback note below.

Callbacks (Webhook)

Provide callback.url when creating a task and HiAPI POSTs to your service once the task reaches a terminal state — no polling needed. The callback payload mirrors the task detail structure; signature verification, delivery retries and idempotency are documented on the Unified Async API overview page.

urlstringRequired

Required when callback is present. A POST is sent to this URL once the task reaches a terminal state; must be a valid http(s) URL.

whenstringOptionalDefault: "final"

Only final is supported today — notified on both success and fail.

Options:final

storagestringOptionalDefault: "temp"

Artifact storage tier: temp (default, retained about 7 days) or persistent (long-term, billed by size). Insufficient balance or exceeding the storage cap silently downgrades to temp; the effective tier is shown in the task detail response.

Options:temppersistent

Request examples

# Submit an async task; returns a taskId immediately (no waiting for generation)
curl -X POST "https://api.hiapi.ai/v1/tasks" \
  -H "Authorization: Bearer $HIAPI_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
  "model": "flux-3-image-klein",
  "input": {
    "steps": 16,
    "prompt": "…",
    "resolution": "1K",
    "aspect_ratio": "auto",
    "output_format": "jpeg",
    "safety_tolerance": 2
  }
}'

const res = await fetch("https://api.hiapi.ai/v1/tasks", {
  method: "POST",
  headers: {
    Authorization: `Bearer ${process.env.HIAPI_API_KEY}`,
    "Content-Type": "application/json",
  },
  body: JSON.stringify({
    model: "flux-3-image-klein",
    input: {
      "steps": 16,
      "prompt": "…",
      "resolution": "1K",
      "aspect_ratio": "auto",
      "output_format": "jpeg",
      "safety_tolerance": 2
    }
  }),
});
const task = await res.json();
// { code: 200, message: "success", data: { taskId } }

import os, requests

res = requests.post(
    "https://api.hiapi.ai/v1/tasks",
    headers={"Authorization": f"Bearer {os.environ['HIAPI_API_KEY']}"},
    json={
        "model": "flux-3-image-klein",
        "input": {
          "steps": 16,
          "prompt": "…",
          "resolution": "1K",
          "aspect_ratio": "auto",
          "output_format": "jpeg",
          "safety_tolerance": 2
        }
    },
)
print(res.json())  # { code, message, data: { taskId } }

200 OK

{
  "code": 200,
  "message": "success",
  "data": {
    "taskId": "tk-hiapi-01HZTQ8BX2N3GM3YFK4Z9D7VQR"
  }
}

400 Bad Request

{
  "code": 400,
  "message": "invalid request",
  "data": null,
  "error_code": "INVALID_REQUEST"
}

Get task detail

GET https://api.hiapi.ai/v1/tasks/:id

Response fields

taskIdstringRequired

Task ID, always prefixed with tk-hiapi-.

modelstringRequired

Resolved full model name. Tasks submitted with the route parameter return base@route (e.g. gpt-image-2/text-to-image@ext); otherwise identical to the name passed at creation.

routestringOptional

Present only when the task was submitted with an explicit route parameter; echoes the submitted value.

statusstringRequired

Task status.

Status values

queued Queued, not started yet. handling Being processed. archiving Processing finished, artifacts being prepared. success Terminal: completed, artifacts available (output returned). fail Terminal: failed (error returned). After submission, check status: success and fail are terminal, everything else is non-terminal. Poll every 1–3 seconds until terminal. On success, download artifacts from output[].url and transfer them before expireAt; pass storage: "persistent" at creation time when long-term retention is needed.

creatednumberRequired

Task creation time (Unix seconds).

completednumberRequired

Unix seconds when the task reached a terminal state; fixed to 0 while not terminal.

outputarrayOptional

Present only when status == success. Artifact list — always an array; a single task may return multiple types at once (e.g. video + first_frame).

Array [ object ]

Array items

urlstringRequired

Artifact URL, directly downloadable.

typestringRequired

Current possible values: video, image, audio, first_frame, last_frame, thumbnail; more may be added later.

expireAtnumberRequired

Unix seconds until which the artifact remains accessible; download or transfer before it expires.

errorobjectOptional

Present only when status == fail.

codestringRequired

Business error code.

messagestringRequired

Error description.

Response examples

Success

{
  "code": 200,
  "message": "success",
  "data": {
    "taskId": "tk-hiapi-01HZTQ8BX2N3GM3YFK4Z9D7VQR",
    "model": "happyhorse-1-0",
    "status": "success",
    "created": 1777282033,
    "completed": 1777282099,
    "output": [
      {
        "url": "https://cdn.hiapi.ai/tasks/.../video.mp4",
        "type": "video",
        "expireAt": 1777886899
      }
    ]
  }
}

Processing (non-terminal)

{
  "code": 200,
  "message": "success",
  "data": {
    "taskId": "tk-hiapi-...",
    "model": "happyhorse-1-0",
    "status": "archiving",
    "created": 1777282033,
    "completed": 0
  }
}

Failed

{
  "code": 200,
  "message": "success",
  "data": {
    "taskId": "tk-hiapi-...",
    "model": "happyhorse-1-0",
    "status": "fail",
    "created": 1777282033,
    "completed": 1777282099,
    "error": {
      "code": "TASK_FAILED",
      "message": "task failed"
    }
  }
}