It accepts a prompt and up to ten reference images, returning one image. The 5.6B count applies to its image transformer; it uses Qwen3-VL-4B-Instruct as a separate text encoder.
The repository contains a distilled checkpoint for generation and editing at 16 steps without guidance, plus BF16, FP8, and NVFP4 variants. The related Klein Base checkpoint is identified for fine-tuning, LoRA training, and research, at 50 steps and guidance 3.5. The model card describes bounding-box placement on a 0–1000 grid, with boxes guiding rather than hard-masking pixels.
Files are gated: users must log in to Hugging Face and accept the repository conditions. The model is under the FLUX Kommunity License, which permits non-commercial use and describes a commercial-output exception for qualifying users. The card says the reference workflow relies on prompt upsampling and warns that short prompts produce generic output. It does not establish comparative image quality, production reliability, or total deployment cost.