FLUX 3 Image Klein: Redefining Open Weights Generation

Date: October 15, 2026 | Read Time: $\approx 4$ minutes Developer: Black Forest Labs

Today, we are proud to introduce FLUX 3 Image Klein, an open weights model featuring 5.6B parameters. This versatile tool is engineered for high-fidelity text-to-image synthesis, precise image manipulation, and multi-reference generation (supporting up to 10 distinct reference images).

---

🚀 Core Capabilities & Architecture

FLUX 3 Image Klein inherits the powerhouse features of the full FLUX 3 Image suite, including:

• Native 4K Resolution: Ultra-high definition output.
• Pixel-Perfect Editing: Surgical precision in image modification.
• Bounding Box Control: Use coordinates to dictate exactly where elements appear or to restrict edits to specific zones.

The Multimodal Synergy

The model's realism stems from shared intelligence across different media types. The integration can be visualized as follows:

---

📊 Performance & Benchmarks

Despite its compact size, FLUX 3 Image Klein outperforms both larger open-source models and proprietary "closed" systems like Ideogram V4.5 and FLUX.2 [pro].

Model Comparison

Efficiency Gains

When compared to the previous FLUX.2 [klein] 9B, the new model achieves:

• $\Delta \text{Elo}_{\text{Text-to-Image}} = +91$
• $\Delta \text{Elo}_{\text{Image Editing}} = +70$
• Parameter Reduction: $\approx 38\%$ fewer parameters.

(Note: Parameter counts refer strictly to the image generator; text encoders are excluded.)

---

💻 Hardware & Deployment

The reduction in size makes this model ideal for cloud-only local deployment and on-device editing on consumer laptops.

Memory Requirements

• BF16 Precision: 11.2 GB
• FP8 Precision: 5.6 GB
• INT8 Build: Compatible with NVIDIA GeForce RTX GPUs with as little as 8GB VRAM.

Generation Speed (1MP Image)

Using the distilled model and INT8 build via ComfyUI:

1. NVIDIA RTX 5090: $\approx 4$ seconds
2. NVIDIA RTX 4070: $< 9$ seconds
3. NVIDIA GB300 (NVFP4): $< 0.6$ seconds

Technical Note: GPU timings represent raw generation. API latency includes network overhead, queueing, and prompt upsampling.

---

🛠️ Ecosystem & Integration

The release provides a base model for fine-tuning and a step distilled model for high-speed production.

Supported Tooling

• Training: LoRA support via Diffusers and Ostris's AI Toolkit; recipes available in NVIDIA NeMo AutoModel.
• Inference Frameworks:
  • ComfyUI
  • Hugging Face Diffusers
  • TensorRT-LLM
  • SGLang Diffusion
  • vLLM-Omni

API Implementation

The distilled model is accessible via the BFL API as flux-3-image-klein. It features:

• Prompt Upsampling
• Pixel Restoration (ensures unedited areas remain bit-identical).

# Example conceptual reference for BFL API
response = bfl_api.generate(
    model="flux-3-image-klein",
    prompt="A futuristic cityscape",
    reference_images=[img1, img2], # Up to 10
    bounding_box={"x": 100, "y": 200, "w": 50, "h": 50}
)

---

⚖️ Licensing & Safety

The FLUX Kommunity License

The weights are released under a tiered structure:

• Free Commercial Use: For individuals and companies with annual revenue < \5\text{M}$.
• Research/Non-Commercial: Free for all.
• Enterprise: Larger organizations require self-hosted licenses or API agreements.

Safety Framework

To prevent misuse, BFL implemented a multi-stage safety pipeline:

• Pre-training: Dataset filters to remove harmful content.
• Post-training: Targeted fine-tuning to suppress malicious outputs.
• Validation: Rigorous internal red-teaming and external auditing by Cinder.

---

📚 Further Reading & Citations

FLUX 3 Image Klein is the second open release in the family, following FLUX 3 Action (the 7B robot control model).

• Research: FLUX 3: Real World Models
• Robot Control: FLUX 3 Action

Citation Format:

@misc{bfl2026flux3imageklein, 
  author = {{Black Forest Labs}}, 
  title = {{FLUX} 3 Image Klein}, 
  year = {2026}
}