Felo API PlatformFelo API Platform
Back to models

FLUX 3 Image API

Black Forest Labs' FLUX 3 Image is an early-access open-weight image generation and editing model with native 4K output, pixel-level control, and multi-reference composition.

Black Forest Labs: FLUX 3 Image is part of the FLUX 3 multimodal foundation model family announced on July 23, 2026. It generates images from text prompts and edits images from an image plus a prompt. The model supports multi-reference generation with up to 10 reference images and produces native 4K resolution output. It offers pixel-level precise editing and uses bounding boxes to control element placement and limit edit regions. Rather than short prompts, it accepts full scene captions, and it pairs with a 9B prompt upsampler model that expands requests into detailed scene descriptions via system prompt.

The image transformer is provided as a 5.6B-parameter open-weight model, with a model size of 6B parameters in BF16 tensor type. It uses Qwen3-VL-4B-Instruct as the text encoder. An undistilled base version (Klein Base) is available for fine-tuning, LoRA training, and research, while a distilled Klein version is offered for generation, editing, and self-hosting. The checkpoint behind the BFL API is the distilled version. It integrates with the Diffusers library from day one and runs on RTX 4070-class GPUs.

FLUX 3 Image is suited to text-to-image generation, single-reference image editing, multi-reference image composition, design and advertising production, and product visualization. According to internal evaluations, FLUX 3 Image Klein placed first in text-to-image, multi-reference generation, and image editing rankings. It improves over FLUX.2 Klein 9B by 91 Elo points in text-to-image and 70 Elo points in image editing, while reducing parameter count by 38%. It outperforms larger open-weight models such as Qwen-Image 2.1 (7B parameters) and Krea 2 (12.9B parameters), and surpasses closed models Ideogram V4.5 and FLUX.2 Pro. Note that the flux-kommunity-license imposes conditions on commercial use.

Updated:

MODEL ID

flux-3-image
Input
text / image
Output
image

Highlights

  • Native 4K image generation
  • Pixel-level precise editing
  • Up to 10 reference images
  • Open-weight 5.6B image transformer
  • Runs on RTX 4070-class GPUs
  • Distilled and base versions

Providers

1 providers

Compare API prices across providers. Select a provider to view all charges and conditions.

Provider · Pricing details
Uptime
Not published

Benchmarks

  • Text-to-image ranking (internal evaluation)

    FLUX 3 Image Klein placed first among compared models

    #1
  • Multi-reference generation ranking (internal evaluation)

    FLUX 3 Image Klein placed first among compared models

    #1
  • Image editing ranking (internal evaluation)

    FLUX 3 Image Klein placed first among compared models

    #1
  • Text-to-image improvement vs FLUX.2 Klein 9B

    Elo points

    91
  • Image editing improvement vs FLUX.2 Klein 9B

    Elo points

    70
  • Physics-IQ

    ranked as the world's strongest world model

    #1

Scores as published by the model provider.

Quick Start

Submit an image task, poll its status, then save the images. These examples use curl and Python 3 (standard library only).

POST https://openapi.felo.ai/api/v1/images/generations

1Get your API key

Create a key from your Felo API Platform account and set it as an environment variable.

export FELO_API_KEY="YOUR_API_KEY"

2Make your first request

Optional values vary by model. This example leaves them unset; confirm supported values in Playground before adding them. Request fields sit alongside model and prompt, without an input wrapper.

curl --fail-with-body --max-time 60 "https://openapi.felo.ai/api/v1/images/generations" \
  -H "Authorization: Bearer $FELO_API_KEY" \
  -H "Content-Type: application/json" \
  -o created.json \
  -d '{
  "prompt": "A studio product photo of a ceramic coffee cup on warm linen, soft morning light",
  "model": "flux-3-image"
}'

3Poll the task

Submission returns task_id. Poll every 3 seconds while queued or running; stop on succeeded, failed or interrupted. This example stops waiting after 30 minutes. Keep created.json to resume querying the same task instead of submitting again.

python3 - <<'PY'
import json, os, time
from pathlib import Path
from urllib.parse import quote
from urllib.request import Request, urlopen

task = json.loads(Path("created.json").read_text())
task_id = task["task_id"]
url = "https://openapi.felo.ai/api/v1/images/generations/" + quote(task_id, safe="")
headers = {"Authorization": "Bearer " + os.environ["FELO_API_KEY"]}
deadline = time.monotonic() + 1800

while task["status"] in ("queued", "running"):
    if time.monotonic() >= deadline:
        raise TimeoutError(f"Keep task_id={task_id} to resume polling")
    time.sleep(3)
    with urlopen(Request(url, headers=headers), timeout=60) as response:
        task = json.load(response)
    if task.get("task_id") != task_id or task.get("kind") != "image":
        raise ValueError("Unexpected task response")
    print(task["status"])

if task["status"] != "succeeded":
    raise RuntimeError(task.get("error") or task["status"])
if not task.get("outputs"):
    raise ValueError("Task succeeded without image outputs")
Path("task.json").write_text(json.dumps(task))
print("Saved task.json")
PY

4Download the output

After success, save outputs from task.json. Results may be HTTPS URLs, URL objects or b64_json. Download signed URLs promptly; media downloads do not use your API key.

python3 - <<'PY'
import base64, json
from pathlib import Path
from urllib.request import urlopen

task = json.loads(Path("task.json").read_text())
if task.get("status") != "succeeded" or not task.get("outputs"):
    raise ValueError("No completed image outputs")
extensions = {"image/png": ".png", "image/jpeg": ".jpg", "image/webp": ".webp", "image/gif": ".gif", "image/avif": ".avif"}

for index, output in enumerate(task["outputs"], 1):
    item = output if isinstance(output, dict) else {}
    url = output if isinstance(output, str) else item.get("url") or item.get("unsigned_url")
    mime = item.get("media_type") or "image/png"
    if url and url.startswith("https://"):
        with urlopen(url, timeout=60) as response:
            data = response.read()
            mime = response.headers.get_content_type()
    elif url and url.startswith("data:image/") and ";base64," in url:
        header, encoded = url.split(",", 1)
        mime = header[5:].split(";", 1)[0]
        data = base64.b64decode(encoded, validate=True)
    elif item.get("b64_json"):
        data = base64.b64decode(item["b64_json"], validate=True)
    else:
        raise ValueError("Unrecognized image output")
    path = Path(f"image-{index}" + extensions.get(mime, ".bin"))
    path.write_bytes(data)
    print(path)
PY
input_references

For image editing, replace the example URL with your HTTPS image URL or a base64 image data URL. Submit once, then use the same polling and saving steps.

curl --fail-with-body --max-time 60 "https://openapi.felo.ai/api/v1/images/generations" \
  -H "Authorization: Bearer $FELO_API_KEY" \
  -H "Content-Type: application/json" \
  -o created.json \
  -d '{
  "prompt": "Keep the product and framing. Replace the background with a sunlit marble countertop.",
  "model": "flux-3-image",
  "input_references": [
    {
      "type": "image_url",
      "image_url": {
        "url": "https://example.com/product.png"
      }
    }
  ]
}'

Supported parameters

  • n
  • resolution
  • aspect_ratio
  • input_references

Prefer the full reference?

API reference →

Frequently asked questions

What is FLUX 3 Image?

Black Forest Labs: FLUX 3 Image is part of the FLUX 3 multimodal foundation model family announced on July 23, 2026. It generates images from text prompts and edits images from an image plus a prompt. The model supports multi-reference generation with up to 10 reference images and produces native 4K resolution output. It offers pixel-level precise editing and uses bounding boxes to control element placement and limit edit regions. Rather than short prompts, it accepts full scene captions, and it pairs with a 9B prompt upsampler model that expands requests into detailed scene descriptions via system prompt. The image transformer is provided as a 5.6B-parameter open-weight model, with a model size of 6B parameters in BF16 tensor type. It uses Qwen3-VL-4B-Instruct as the text encoder. An undistilled base version (Klein Base) is available for fine-tuning, LoRA training, and research, while a distilled Klein version is offered for generation, editing, and self-hosting. The checkpoint behind the BFL API is the distilled version. It integrates with the Diffusers library from day one and runs on RTX 4070-class GPUs. FLUX 3 Image is suited to text-to-image generation, single-reference image editing, multi-reference image composition, design and advertising production, and product visualization. According to internal evaluations, FLUX 3 Image Klein placed first in text-to-image, multi-reference generation, and image editing rankings. It improves over FLUX.2 Klein 9B by 91 Elo points in text-to-image and 70 Elo points in image editing, while reducing parameter count by 38%. It outperforms larger open-weight models such as Qwen-Image 2.1 (7B parameters) and Krea 2 (12.9B parameters), and surpasses closed models Ideogram V4.5 and FLUX.2 Pro. Note that the flux-kommunity-license imposes conditions on commercial use.

How is usage billed?

Your bill depends on the provider you use, its prices and your usage. Longer inputs or other conditions may have different prices.

What inputs does FLUX 3 Image accept?

It accepts text input and image input, and produces image output. It supports text-to-image generation and image editing from an image plus a prompt, including multi-reference generation with up to 10 reference images.

What resolution does FLUX 3 Image generate at?

It generates at native 4K resolution.

What editing control does FLUX 3 Image provide?

It provides pixel-level precise editing and uses bounding boxes to control element placement and limit edit regions.

What is the difference between the base and distilled versions?

The undistilled base version (Klein Base) is intended for fine-tuning, LoRA training, and research. The distilled Klein version is offered for generation, editing, and self-hosting. The base version runs at 50 steps with guidance 3.5, while the distilled version runs at 16 steps without guidance.

Does FLUX 3 Image require a separate model for prompts?

Yes. It pairs with a 9B prompt upsampler model that expands requests into detailed scene descriptions via system prompt.

What are the licensing considerations?

The flux-kommunity-license imposes conditions on commercial use. The Qwen3-VL-4B-Instruct text encoder is under the Apache-2.0 license.

Related models