Black Forest Labs Releases FLUX.3 Image Klein in Diffusers

FLUX.3 Image Klein Hugging Face Diffusers Release

Black Forest Labs has brought its newest open image generation model, FLUX.3 Image Klein, to the Hugging Face Diffusers library on the same day as its public debut. The release gives developers a single model that handles text-to-image generation, multi-reference image-to-image editing, and a fresh layout control feature that lets users place specific elements by writing bounding boxes directly into the prompt. Both model checkpoints are already hosted in the FLUX.3 Image Klein collection on the Hugging Face Hub, and a public Space lets anyone experiment with the model immediately.

The lineup includes two variants. The primary checkpoint, black-forest-labs/flux-3-image-klein, is a time-step distilled version optimized for fast inference. It produces a 1024 by 1024 pixel image in sixteen steps without classifier-free guidance, completing the job in roughly two seconds on an H100 GPU. The second checkpoint, black-forest-labs/flux-3-image-klein-base, is the full base model that runs fifty steps with classifier-free guidance for users who prefer the standard diffusion trajectory. For teams working with smaller GPUs, NVFP4 and FP8 quantized weights are available and can be loaded through the from_single_file method, lowering the memory barrier considerably.

FLUX.3 Image Klein Installation Requirements

Getting started requires only the latest Diffusers and Transformers packages, the latter needed for the Qwen3-VL text encoder that the model uses. The pipeline loads through the new ModularPipeline class, which reflects a broader architectural shift in the library. Instead of a monolithic object, the pipeline is assembled from interchangeable components. That modularity means developers can swap out individual blocks, insert custom steps inside the denoising loop, or encode a prompt once and reuse the embeddings across multiple generations without recomputing them.

FLUX.3 Image Klein Layout Control Multi-Reference

The layout control capability is the most distinctive addition. Users append a JSON array of region objects to their prompt, each carrying an identifier, a bounding box, and a natural-language description of what should appear inside that box. The boxes use a normalized 0 to 1000 scale with the y coordinate first, so the same coordinates work at any output resolution. The prompt then references each region by its identifier using angle-bracket syntax. This lets creators compose scenes by specifying, for example, that a chartreuse background should fill the entire frame while a stippled-ink silhouette occupies a centered square. The same mechanism works for editing existing images. By providing a reference image and specifying both a source bounding box and a target bounding box for a region, the model can move an object, such as a crocheted knight figure, to a new position while pinning the surrounding backdrop, cliff, dragon, and other elements exactly where they were.

Multi-reference image-to-image works by passing a list of images to the pipeline and referring to them in the prompt as ref_image_0, ref_image_1, and so on. A prompt can then describe a composition that pulls the subject from one reference and the setting from another, letting the model fuse them into a coherent new scene. The model responds best to detailed scene descriptions that cover subject, setting, lighting, composition, and style, following the prompting guide published by Black Forest Labs.

FLUX.3 Image Klein LoRA Fine-Tuning Scripts

On the training side, the release includes LoRA fine-tuning scripts tailored for both text-to-image and image-editing workflows. That gives teams a path to adapt the model to specific visual styles or domain-specific content without full retraining. The modular pipeline design also makes it easier to experiment with architectural changes, such as inserting a custom attention block or replacing the scheduler, because each piece is exposed as a separate component.

For builders, the immediate takeaway is that a state-of-the-art open image model with novel spatial control is now part of the standard Diffusers ecosystem. The distilled checkpoint offers a practical speed-memory trade-off for production inference, while the base model provides a foundation for research and fine-tuning. The layout control syntax, though new, is straightforward enough to integrate into existing prompt-engineering pipelines, and the modular architecture reduces the friction of customizing the generation loop. Anyone with a CUDA environment can install the updated packages and start generating images today, while teams planning to deploy at scale should evaluate the quantized variants against their hardware constraints.

Source: Hugging Face