FLUX 3 Image Klein packs 4K generation into a 5.6B model

FLUX 3 Image Klein is an open-weights 5.6B parameter model by Black Forest Labs built for fast text-to-image generation, pixel-perfect editing, and multi-reference workflows with up to 10 images, all supporting native 4K output.

Its standout feature is built-in bounding box layout control. You can pass normalized coordinates right alongside your prompt to position specific subjects, or define source and target bounding boxes to move elements directly from a reference image.

How it works and hardware requirements

The model builds on FLUX 3's shared cross-modal training across images, video, and audio to ground generated details in real-world physical dynamics.

It arrives in two checkpoints: a base model for fine-tuning and a step distilled model for low-latency generation. Black Forest Labs partnered with NVIDIA to provide FP8, INT8, and NVFP4 formats right out of the gate.

Hardware requirements are remarkably light. Weights take 11.2GB in BF16 and 5.6GB in FP8. With the INT8 build, it runs comfortably on consumer NVIDIA GeForce RTX cards with as little as 8GB VRAM through ComfyUI or Diffusers.

In terms of speed, the distilled INT8 model produces a 1MP image in about 4 seconds on an RTX 5090 and under 9 seconds on an RTX 4070. On an NVIDIA GB300 in NVFP4, that drops below 0.6 seconds.

Performance and early reception

In internal benchmarks, Black Forest Labs ranks it ahead of larger open models like Qwen-Image 2.1 (7B) and Krea 2 (12.9B), and even above closed models like Ideogram V4.5. Compared to FLUX.2 Klein (9B), it scores 91 Elo higher in text-to-image with 38% fewer parameters.

However, early community impressions are much more mixed. Some testers running it locally in ComfyUI reported severe anatomy and face deformations, comparing the broken human figures and extra fingers to the early days of Stable Diffusion 3.

The tooling ecosystem is already ready. Diffusers supports it via a modular pipeline paired with a Qwen3-VL text encoder, and Day 1 integration is live across ComfyUI, TensorRT-LLM, vLLM-Omni, and SGLang Diffusion.

Access and licensing

Weights are available on Hugging Face alongside inference scripts and Gradio spaces. LoRA training is supported on Day 1 using Ostris's AI Toolkit and NVIDIA NeMo AutoModel recipes.

It is released under the FLUX Kommunity License, granting free commercial use to individuals and companies making under $5M in annual revenue. For cloud users, the BFL API provides the distilled model starting at $0.009 per 1,000 images.

Why it matters

We think getting 4K generation, multi-reference conditioning, and native bounding box controls into an 8GB VRAM envelope is a genuine engineering win for indie developers and local runners.

That said, ordinary creators looking for polished photorealism should probably wait. The initial reports of anatomy horror show that shrinking to 5.6B came with compromises, and it will take community fine-tunes and LoRAs to make this reliable for daily production.