Black Forest Labs' FLUX 3 Klein Runs Image Generation on 8GB GPUs

Black Forest Labs has released FLUX 3 Image Klein, a 5.6-billion-parameter open-weights model for text-to-image generation, image editing, and multi-reference composition. It follows FLUX 3 Action as the second open release in the FLUX 3 family.

The model targets developers who need local deployment and fine-tuning without the hardware demands of larger image generators. Its weights occupy 11.2GB in BF16 and 5.6GB in FP8. An INT8 build can run through ComfyUI or Diffusers on NVIDIA GeForce RTX GPUs with 8GB of VRAM.

A small generator with a larger toolkit

Black Forest Labs released a base checkpoint for fine-tuning, a step-distilled checkpoint that requires fewer inference steps, and quantized builds produced with NVIDIA. The repository rounds the base model to 6B parameters, while the reported parameter count is 5.6B.

The launch includes a ComfyUI build, a Gradio demo with bounding-box controls, and support in Diffusers, TensorRT-LLM, SGLang Diffusion, and vLLM-Omni.

The weights use the FLUX Kommunity License. Individuals and companies with less than $5 million in annual revenue receive free commercial use alongside research and noncommercial rights. Larger organizations need a self-hosted commercial license or the hosted API, whose pricing starts at $0.009 for a 1K-resolution image. The license terms still govern deployment and redistribution despite the availability of the weights.

Benchmarks favor Klein, with caveats

Black Forest Labs reports that Klein ranks first among the open-weights image models in its third-party evaluation set. The company’s internal tests also place it first for text-to-image generation, multi-reference composition, and image editing.

The comparison includes the closed Ideogram V4.5 and FLUX.2 pro systems, both of which receive lower reported ratings. Klein is also the smallest open model in the group at 5.6B parameters, compared with 7B for Qwen-Image 2.1 and 12.9B for Krea 2.

Against the lab’s earlier FLUX.2 klein 9B model, the new release gains 91 Elo points in text-to-image generation and 70 points in image editing while using 38 percent fewer parameters. Elo summarizes pairwise evaluator preferences, so the gains indicate relative ranking rather than a percentage improvement in image quality.

These comparisons remain vendor-reported even when the underlying evaluations come from third parties. Prompt sets, inference settings, evaluator design, and confidence intervals determine how well the rankings transfer to a specific workload.

Consumer GPUs reach practical speeds

Black Forest Labs reports the following latency for one-megapixel generation with the distilled model:

The GeForce measurements use ComfyUI, while the GB300 result targets datacenter deployment. Actual latency will vary with resolution, inference steps, attention implementation, memory transfers, and whether the prompt-upsampling model runs in the same request path.

The company also reports lower hosted-API latency than Qwen-Image 2.1, FLUX.2 pro, Ideogram V4.5, and Krea 2. Within that comparison, no model at the same or greater latency received a higher quality rating.

Editing controls protect untouched regions

Klein focuses on changing the region described by an edit prompt while preserving the surrounding image. Diffusion-based editors often introduce small changes outside the target area, including shifts in texture, lighting, or facial details.

The hosted API addresses that problem with pixel restoration, which Black Forest Labs says preserves unedited regions exactly. Bounding boxes provide spatial constraints for edits and new objects, and the reference Gradio application exposes those controls directly. Pixel restoration is identified as an API feature, so developers using raw checkpoints should evaluate local editing behavior separately.

Multi-reference generation accepts as many as 10 input images, supporting workflows that combine character identity, products, visual style, and layout. The model also inherits native 4K generation from the full FLUX 3 Image system, though higher-resolution output will require more memory and time than the published one-megapixel latency tests.

Distillation carries over FLUX 3 training

Klein is distilled from the full FLUX 3 Image model, allowing the smaller network to learn from the larger system’s outputs and complete generation in fewer steps. The base checkpoint preserves a more flexible starting point for adaptation, while the distilled checkpoint prioritizes inference speed.

Black Forest Labs says the FLUX 3 family shares training signals across images, video, and audio. Images provide spatial detail, video contributes motion and physical dynamics, and audio adds information about events and interactions. The company uses this multimodal training to improve representations of objects and physical relationships.

Training ran on NVIDIA GB200 NVL72 systems. Deployment optimizations target TensorRT-LLM, vLLM-Omni, and SGLang Diffusion, while the base checkpoint supports LoRA fine-tuning through Ostris’s AI Toolkit and NVIDIA NeMo AutoModel recipes. LoRA trains compact adapter weights, reducing the compute and storage required for customization.

Deployment choices change the footprint

Prompt upsampling is part of the recommended generation pipeline. The hosted API applies it automatically, while local deployments need the supplied system prompt and the separate flux-3-image-klein-pup model. Because that helper has 9B parameters, teams must budget for it separately through memory allocation, offloading, or sequential execution.

A practical deployment sequence is:

1. Choose the base checkpoint for fine-tuning or the distilled checkpoint for faster inference.
2. Select BF16, FP8, INT8, or NVFP4 according to the target GPU and runtime.
3. Add the recommended prompt-upsampling stage when output quality takes priority.
4. Use bounding boxes when object placement or edit boundaries must remain predictable.
5. Measure latency and memory with the intended resolution, reference count, and concurrency.
6. Confirm that the planned commercial use fits the license terms.

The release materials are available through the following resources:

• Model collection for checkpoints and quantized builds
• Live demo for browser-based testing
• API docs for hosted generation

Where Klein fits

Klein suits teams that need access to model weights, consumer-GPU inference, or application-specific adaptation. Likely uses include:

• On-device and desktop photo-editing tools
• LoRA customization for brands, products, or recurring characters
• Region-constrained editing inside design software
• Multi-reference composition for product and marketing assets
• Research that requires checkpoints for ablations and fine-tuning

Teams that prioritize managed infrastructure and occasional generation can use the hosted API. Workloads centered on maximum output quality, with fewer constraints on latency and deployment control, may still favor the full FLUX 3 Image model or another hosted system.