Flux 3 Image Klein

Flux 3 Image Klein is a distilled, open-weights image transformer package (5.6B parameters for the image model, with a separate Qwen3-VL-4B-Instruct text encoder) that performs text-to-image generation, single-reference image editing and multi-reference composition using the same model. It accepts a text prompt plus up to ten reference images and returns a single generated image; a separate “Base” checkpoint exists for fine-tuning and LoRA training. The repo supplies the BF16 distilled checkpoint, quantized FP8 and NVFP4 variants, an image autoencoder, and the text encoder, and provides Diffusers and reference-CLI examples for generation and editing. A prompt upsampler model is available to expand short user instructions into full scene captions; Diffusers can load the modular pipeline and the upsampler can be run via Transformers.

Key capabilities include layout control via JSON-formatted bounding boxes on a 0-1000 grid, in-image text rendering, and compositional edits that use reference images as anchors rather than hard masks. Practical notes: quantized packages require compatible loaders, the prompt upsampler is large (~18.8 GB) and can add ~28 GB GPU usage, and peak VRAM with reference code is ~31 GB (BF16), 26 GB (FP8) or 25 GB (NVFP4) at 1024×1024. Limitations include inaccurate or distorted rendered text, bias amplification, imperfect prompt following, elements bleeding beyond box bounds, and edits affecting areas outside target regions. Distribution uses the FLUX Kommunity License (non-commercial with a commercial-output exception for qualifying users).