FLUX 3 Image Klein brings 5.6B open weights and multi-reference editing

ComfyUI, the node-based generative AI workflow interface, added native support for FLUX 3 Image Klein, a 5.6B parameter model released with open weights. This integration allows users to run both the base and distilled versions of the model directly within their local environments, providing a measurable increase in spatial control and multi-image referencing. The update shifts the model from a cloud-only dependency to a local tool for creators requiring high precision in their generative pipelines.

What's new

FLUX 3 Image Klein introduces a 5.6B parameter architecture that balances performance with hardware accessibility. The release includes two distinct versions: the standard base weights and a distilled version optimized for faster inference. As of February 2025, these models are available as open weights, allowing for local deployment without reliance on proprietary cloud APIs. This makes the model particularly useful for studios that require data privacy or offline functionality.

The model's main technical addition is its native support for bounding boxes and multi-reference editing. Unlike previous iterations that required complex ControlNet stacks or IP-Adapter setups, FLUX 3 Image Klein handles spatial positioning and style referencing within the core model architecture. This allows ComfyUI users to define specific areas for objects using coordinates and feed multiple reference images to guide the aesthetic or structural output of a single generation. The implementation in ComfyUI utilizes a set of dedicated nodes designed to handle these multi-reference inputs, weighing the influence of each reference image against the text prompt to prevent concept bleeding.

How it fits your workflow

For technical artists and workflow designers, FLUX 3 Image Klein provides a middle ground between the massive parameter counts of Flux 1.1 Pro and the lighter Stable Diffusion XL models. It serves as a direct alternative to Stable Diffusion 3.5 Large, offering comparable prompt adherence but with superior handling of multi-subject compositions through its bounding box logic. This is a practical choice for professional storyboard artists and concept designers who require exact placement of assets within a frame.

In a ComfyUI environment, this model simplifies the creation of complex image-to-image pipelines. Editors can now use the Image Klein nodes to reference a character's face from one image and their clothing from another, while placing them in a specific quadrant of the frame using bounding box inputs. This level of control is often difficult to achieve in tools like Midjourney or basic DALL-E 3 interfaces. Compared to the previous Flux 1.0 Dev model, FLUX 3 Image Klein shows improved efficiency in VRAM usage, making it more viable for users running consumer-grade GPUs with 12GB to 16GB of memory. The distilled weights specifically target faster iteration cycles, which is useful for VFX artists using ComfyUI for rapid prototyping or texture generation.

What it costs / how to try it

FLUX 3 Image Klein is available as an open-weights release, meaning it is free to download and use for local ComfyUI installations. Users can update their ComfyUI instance and download the model checkpoints via the ComfyUI Manager or directly from the official repository as of February 2025.