Black Forest Labs' FLUX.2-klein-9B, a 9-billion-parameter model that generates images in under one second, now has updated ComfyUI integration through the NKD Klein Tools v1.7.0 nodes, released this week with a new megapixels slider and image fit control.

What Happened

FLUX.2-klein-9B is a rectified flow transformer from Black Forest Labs that runs at sub-second latency on an RTX 4090, using only 4 inference steps. The model processes images directly in pixel space using a presampling and postsampling pipeline, rather than the compressed latent space used by earlier FLUX.1 models.

The NKD Klein Tools project wraps this pipeline into two ComfyUI V3 nodes: NKD Klein Presampling and NKD Klein Postsampling, which auto-detect the operating mode based on inputs: text-to-image, image-to-image, inpainting, or detail enhancement. Version 1.7.0 replaces the fixed megapixels dropdown with a continuous slider ranging from 0.1MP to 4.0MP, and adds Image Fit options (Native, Center Crop, Outpaint) for handling aspect ratio differences without distortion.

Why It Matters for Creators

Speed is the constraint that shapes iterative creative work. At 15 to 30 seconds per generation with earlier FLUX.1 models, refining a prompt requires waiting through each cycle. FLUX.2-klein-9B generating at sub-second latency changes that into a real-time feedback loop, which is a fundamentally different creative workflow. The model card on HuggingFace reports that FLUX.2-klein-9B matches or exceeds models five times its size in quality benchmarks while requiring approximately 29GB of VRAM, putting it within range of RTX 4090 cards with CPU offloading.

The unified architecture supports text-to-image, image-to-image, inpainting, and detail enhancement within a single model, reducing the overhead of managing separate model checkpoints for different tasks in the same session.

Key Details

  • Parameters: 9 billion (includes 8B Qwen3 text embedder)
  • Inference steps: 4 (step-distilled for speed)
  • Speed: Sub-second end-to-end on RTX 4090
  • VRAM: About 29GB with CPU offloading enabled
  • Modes: Text-to-image, image-to-image, inpainting, detail enhancement
  • License: FLUX Non-Commercial License (research and personal use)
  • New in NKD Klein Tools v1.7.0: Megapixels slider (0.1 to 4.0MP), Image Fit control (Native, Center Crop, Outpaint)

What to Do Next

Install NKD Klein Tools from ComfyUI Manager by searching for "NKD Klein Tools," or clone directly into your ComfyUI/custom_nodes/ folder. You also need to download the FLUX.2-klein-9B weights from HuggingFace and accept the non-commercial license terms before the nodes can load the model.

For context on how quickly the ComfyUI node ecosystem is expanding, see our coverage of ComfyUI 0.21.1 adding Claude, Grok, and OpenAI image nodes.

Creator Outcome: Replace 20-second generation waits with real-time image feedback in your ComfyUI workflow. FLUX.2-klein-9B's 4-step pipeline on an RTX 4090 lets you iterate on prompt and reference image combinations interactively rather than in batches. The new megapixels slider in v1.7.0 lets you dial output resolution directly instead of managing width and height separately.