FLUX 3 Image Klein: Features, Benchmarks, Pricing and API Guide

Black Forest Labs released FLUX 3 Image Klein on October 15, 2026. It is a 5.6B-parameter open-weights model for text-to-image generation, image editing and multi-reference composition with up to 10 reference images. It is the smaller member of the FLUX 3 Image family and the first FLUX 3 image model with downloadable weights.

FLUX 3 Image Klein is now available on PoYo through two model IDs, flux-3-klein and flux-3-klein-edit. The FLUX 3 Image Klein API page includes the live generator, supported controls and pricing.

This guide covers what Klein is, what BFL claims about its quality, how it differs from FLUX.2 [klein] and the full FLUX 3 Image model, and how to call it through the PoYo API.

FLUX 3 Image Klein at a Glance

What Is FLUX 3 Image Klein?

FLUX 3 Image Klein is a compact, open-weights version of FLUX 3 Image. According to BFL's release notes, it supports the same core features as the full model: text-to-image, image editing, multi-reference generation and bounding boxes.

The word "Klein" means "small" in German. BFL used the same naming for FLUX.2 [klein], its earlier compact model. The new version builds on the FLUX 3 foundation, which was trained jointly on images, video and audio. BFL says the video and audio signals give the image model a better sense of motion, physical dynamics and real-world interactions.

One model handles every image task. There is no separate editing model inside the weights: you send a prompt alone to generate, or a prompt with reference images to edit or compose.

Two checkpoints

BFL ships two versions of the weights:

• Klein (distilled): 16 sampling steps, no classifier-free guidance. Built for fast generation and self-hosting. This is the checkpoint behind the BFL API.
• Klein Base (undistilled): 50 steps with guidance. Built for fine-tuning, LoRA training and research. LoRAs trained on Base also load into the distilled model.

Key Features

Text-to-image with native 4K

Klein generates images up to 4K. Hugging Face measured the distilled checkpoint at about 2 seconds for a 1024 × 1024 image on an H100.

Multi-reference editing

You can pass up to 10 reference images and refer to them in the prompt by position. This is useful for keeping a product, character or style consistent across a set of images without training a LoRA.

Bounding-box layout control

This is the new capability in the FLUX 3 Image family. You append a JSON list of regions to the prompt. Each region has an ID, a box on a 0–1000 grid and a description. For edits, a region can specify a source box in the reference and a target box in the output, which lets you move, resize, recolor or remove elements.

BFL notes that boxes guide placement and scale but are not hard pixel masks. An element can extend past its box, and very small requested elements may not appear.

Text rendering

Klein can render text inside images, such as product labels, posters and signs. BFL's model card still warns that rendered text can be inaccurate, so check every string before publishing.

Prompt upsampling

Klein is trained on full scene captions. BFL says short prompts produce generic results and recommends a prompt upsampler that expands a request into a detailed caption covering style, scene, lighting, camera and each subject. BFL publishes a 9B upsampler model for self-hosted setups.

In practice, write prompts that describe the subject, setting, light, composition and style. For edits, describe the full result with the change included rather than writing a one-line instruction like "make it blue".

FLUX 3 Image Klein Benchmarks

BFL published internal human-preference evaluations at launch. The company reports that Klein:

• Ranks first among the compared open-weights models in text-to-image, multi-reference generation and image editing
• Rates above the closed models in its comparison, Ideogram V4.5 and FLUX.2 [pro]
• Is the smallest open model in the comparison (5.6B, against 7B for Qwen-Image 2.1 and 12.9B for Krea 2)
• Gains 91 Elo in text-to-image and 70 Elo in image editing over FLUX.2 [klein] 9B, with 38% fewer parameters

These are vendor-published results. Early independent tests are more mixed. One third-party review found Klein followed detailed multi-part instructions less reliably than larger closed models such as Nano Banana 2.1 and GPT Image 2.5 Flare. Klein's advantage is quality per parameter, latency and cost, not winning every complex prompt.

FLUX 3 Image Klein vs FLUX.2 [klein]

If you already use FLUX.2 [klein] for self-hosting, Klein is the direct upgrade: smaller, faster and better on BFL's evaluations, with layout control added.

FLUX 3 Image Klein vs FLUX 3 Image

Both models share the same request format and features. The difference is size, price and access.

At 1K, Klein costs about a fifth of the full model. At 4K, the gap is much wider. For catalog images, social content and drafts, Klein is usually the practical choice. Reserve the full model for hero assets where the extra quality matters.

Hardware and Local Use

BFL says the generator weights need 11.2 GB in BF16 and 5.6 GB in FP8. With the INT8 build, Klein runs on NVIDIA GeForce RTX GPUs with as little as 8 GB of VRAM through ComfyUI or Diffusers.

NVIDIA-measured timings for the distilled model with INT8 are about 4 seconds per 1MP image on an RTX 5090 and under 9 seconds on an RTX 4070. Note that the optional prompt upsampler is a separate 9B model that needs its own memory.

Klein has day-one support in ComfyUI, Hugging Face Diffusers, TensorRT-LLM, SGLang Diffusion and vLLM-Omni. LoRA training works with Ostris's AI Toolkit and Diffusers.

License: Can You Use FLUX 3 Image Klein Commercially?

The weights use the FLUX Kommunity License. It allows research and non-commercial use, plus commercial use of outputs for individuals and companies with less than $5M in annual revenue. Larger organizations need a self-hosted commercial license or an API.

Using Klein through an API such as PoYo avoids self-hosting license questions for the model weights. Always review the provider's terms for your use case.

FLUX 3 Image Klein Pricing on PoYo

PoYo charges per image by resolution. Text-to-image and editing cost the same.

You can request up to four images per call with n. The total cost is the per-image price multiplied by n. A batch of four 2K images costs $0.056.

How to Use the FLUX 3 Image Klein API on PoYo

PoYo exposes two model IDs:

• flux-3-klein for text-to-image. Send a prompt only.
• flux-3-klein-edit for editing and multi-reference generation. Send a prompt and 1–10 images in image_urls.

Submit a task to POST /api/generate/submit, then poll GET /api/generate/status/{task_id} or set a callback_url.

{
  "model": "flux-3-klein-edit",
  "input": {
    "prompt": "The ceramic mug from the first image on a walnut desk beside the notebook from the second image, soft morning window light from the left, 50mm product photo, shallow depth of field, clean minimal background.",
    "image_urls": [
      "https://example.com/mug.png",
      "https://example.com/notebook.png"
    ],
    "size": "4:3",
    "resolution": "2K",
    "n": 2,
    "output_format": "webp"
  }
}

Supported parameters

• size: 1:1, 4:3, 3:4, 16:9, 9:16, 3:2, 2:3, or a custom WIDTHxHEIGHT. Custom sizes must be 256–4096 pixels per side, in multiples of 16, and match the selected resolution.
• resolution: 1K, 2K or 4K. The default is 1K.
• n: 1–4 images per request.
• output_format: png, jpeg or webp.
• enable_safety_checker: on by default.

See the FLUX 3 Image Klein API documentation for the full request schema and response format.

Best Use Cases

• E-commerce: place products in new scenes while keeping the item consistent across references.
• Marketing and brand assets: generate campaign variations in several aspect ratios at low cost.
• Social content: produce high volumes of drafts at 1K, then upscale the winners to 4K.
• Interior and real estate: restyle rooms or stage spaces from a reference photo.
• Storyboards and publishing: keep characters consistent across panels with multi-reference prompts.

Frequently Asked Questions

What is FLUX 3 Image Klein?

It is a 5.6B-parameter open-weights image generation and editing model from Black Forest Labs, released on October 15, 2026. It supports text-to-image, editing, up to 10 references and bounding-box layout control.

Is FLUX 3 Image Klein open source?

It is open weights, not open source in the OSI sense. The weights are on Hugging Face under the FLUX Kommunity License, which limits free commercial use to companies with less than $5M in annual revenue.

How much does FLUX 3 Image Klein cost?

On PoYo, it costs $0.009 per 1K image, $0.014 per 2K image and $0.036 per 4K image. Editing costs the same as generation.

What GPU do I need to run FLUX 3 Image Klein locally?

BFL says the INT8 build runs on NVIDIA GeForce RTX GPUs with 8 GB of VRAM. BF16 weights need about 11.2 GB, and the optional prompt upsampler needs additional memory.

How is FLUX 3 Image Klein different from FLUX 3?

FLUX 3 is the multimodal foundation that powers both image and video models. On PoYo, FLUX 3 refers to the video workflows, while Klein is a compact image model built on the same foundation. See the FLUX 3 overview for the full family.

Is FLUX 3 Image Klein better than FLUX.2?

On BFL's internal evaluations, it beats FLUX.2 [klein] 9B by 91 Elo in text-to-image and 70 Elo in editing, and rates above FLUX.2 [pro]. If you need the existing FLUX.2 Pro and Flex endpoints, they remain available on the FLUX 2 page.

Final Takeaway

FLUX 3 Image Klein brings most of the FLUX 3 Image feature set into a model small enough to run on a consumer GPU and cheap enough for high-volume API use. Its strengths are speed, cost, multi-reference editing and the new bounding-box layout control. Its limits are typical of a compact model: write detailed prompts, and check text rendering and complex instructions before production.

Try it on the PoYo FLUX 3 Image Klein page, or read the API documentation to start integrating.

Sources: Black Forest Labs FLUX 3 Image Klein announcement, FLUX 3 Image Klein model card, BFL release notes, Hugging Face Diffusers blog