Flux 2 Klein Install and Run Locally

No disrespect to GLM image or Qwen image, but when Flux gets released it stays released. The realism here is strong, with vivid images generated either from a text prompt or by amalgamating different images.

This is a new model from Black Forest Labs: Flux 2 Klein 9 billion. It represents a genuine inflection point in generative AI. It doesn't just claim to be fast and good, it actually delivers on both through fundamental architectural innovation. I installed it locally and tested it on various benchmarks and prompts to see if that holds up.

Flux 2 Klein Install and Run Locally

Model at a glance

• 9 billion parameter rectified flow transformer, step distilled down to just four inference steps.
• Variants include a 4 billion model.
• Enables subsecond image generation on consumer hardware like RTX 4090.
• Matches or exceeds models five times its size while running in under half a second, based on benchmarks.
• Unifies text-to-image generation and multi-reference editing in a single architecture, so you are not switching between different systems if you are creating from scratch or modifying existing images.

System setup and prerequisites

I used an Ubuntu system. I started with an Nvidia RTX 6000 with 48 GB of VRAM. The model card says it takes around 29 GB VRAM, yet I still hit out-of-memory after the initial download stage. I upgraded to a new server with an Nvidia H100 and reran everything. VRAM consumption on heavy runs went up to around 72 GB, but the speed was excellent.

Installation steps

1. Install the Torch stack.
2. Clone the Flux tool repository.
3. From the scripts directory, run the app script. I modified their CLI to app.py to expose a Gradio interface.
4. On first run, the script downloads the model.
5. This is a gated model. Go to Hugging Face, log in, and accept the terms and conditions.

Running locally and performance notes

• The Gradio demo ran smoothly in the browser with Flux 2 Klein 9B.
• Speed is the standout - very fast generations with the distilled 4-step setting.
• VRAM usage can be high on the base model and long-step runs, peaking around 72 GB in my tests.

Flux 2 Klein Install and Run Locally - Benchmarks and Prompts

Text-to-image: realism and speed

• Prompt: a weathered fisherman's hand mending a net at golden hour, shallow depth of field, salt crystals on skin, worn rope texture, backlit fingers.
  • Result: Speed was great. The quality, golden hour lighting, visible crystals on skin, and worn rope texture all matched the prompt strongly.

• Prompt: rain-soaked Tokyo alleyway at night, neon reflections in puddles, steam rising from manhole.
  • Result: Very fast. Instruction following was excellent, with crisp reflections and atmosphere.

Materials and textures

• Prompt: extreme close-up of a glass of whiskey with a single ice cube, condensation droplets on the glass, scratched wooden bar surface.
  • Result: Condensation droplets and surface textures looked excellent.

Complex scene composition

• Prompt: a busy mechanics workshop, tools scattered on the workbench.
  • Result: Strong realism. Tools and shapes held up without malformation. Lighting and shadows, including a tube light, looked convincing.

Challenging subjects

• Prompt: a photojournalistic shot of a street musician, worn leather boot tapping rhythm, motion blur.
  • Result: The scene conveyed the subject clearly through details like the boot, coins, and stand. Motion and context read well.

Image editing and multi-reference behavior

• Input image edit: place the scene during heavy rainfall with wet surfaces and reflections.
  • Distilled 4-step result: Okay but not at the level of the best text-to-image outputs.
  • Base 50-step result: Took longer and looked worse than the distilled output in this case.
  • VRAM during 50 steps: around 72 GB.

• Edit: replace banana with a wooden sign reading "a revolution starts here", rustic handcrafted look.
  • Result: The distilled output was better. Spellings were good, which is typically solid with Flux.

• Edits on a portrait:
  • Change lipstick color to black: worked well.
  • Change t-shirt color to red: worked well.
  • Replace t-shirt with a business shirt: worked well and stayed safe.
  • Change eye color to blue: introduced an AI look that did not feel as natural.

Overall, editing is good with targeted changes, but text-to-image is the standout.

Distilled vs base comparison on technical photography

• Prompt: technical fitness photography - lens and lighting control, realistic physical characteristics like perspiration and skin texture, gym setting, natural lighting, strong composition.
  • Distilled result: Quite good.
  • Base result with more steps: In this prompt, the distilled image looked better. The base showed a plasticky look. I asked for perspiration - the distilled output delivered more convincingly. In many AI images you still see some plasticky look one way or another, but this is much improved.

Flux 2 Klein Install and Run Locally - VRAM and Speed

• Distilled 4-step runs: remarkably fast, often subsecond on strong consumer GPUs.
• Base high-step runs: can be heavy. I observed around 72 GB VRAM usage during a 50-step test.
• I hit out-of-memory on a 48 GB RTX 6000 during initial attempts, even though the model card listed ~29 GB, and then moved to a larger GPU.

Final Thoughts

Flux 2 Klein 9B is fast and produces high-quality images, especially in text-to-image generation. Instruction following is strong, materials and complex scenes hold up, and speed is a clear highlight. Image editing works well for many targeted modifications, though some cases introduce an AI look. In several head-to-head prompts, the distilled 4-step model produced results I preferred over the base with more steps.

If you have the VRAM headroom, the 9B distilled model is the go-to. Subsecond generations with realistic lighting, textures, and composition make it a compelling local setup.