flux1-kontext-dev-fp8

FP8-quantized weights for FLUX.1-Kontext-dev diffusion models. Supports both E4M3FN and E5M2 formats.

Model Overview

• Base Model: FLUX.1-Kontext-dev (diffusion model component)
• Quantization: Per-tensor dynamic quantization to FP8 (E4M3FN/E5M2)
• Size Reduction: ~40% smaller than original weights
• Model Scope: This is only the diffusion_model component (not full pipeline), should be placed in ComfyUI's diffusion_models directory
• Compatibility:
  • ✅ ComfyUI (with PyTorch 2.4+ and CUDA 12.4+)
  • ❌ Diffusers (currently unsupported)

Available Files

• flux1-kontext-dev-fp8-e4m3fn.safetensors - Balanced performance/accuracy
• flux1-kontext-dev-fp8-e5m2.safetensors - Higher throughput on Hopper GPUs

Usage

1. Move the .safetensors file to your ComfyUI diffusion_models directory
2. Select fp8 modes in UNet Loader like fp8_e4m3fn.

Known Limitations

1. Not compatible with standard Diffusers pipelines
2. Requires patched PyTorch versions for optimal performance

License

This model is distributed under the FLUX.1(dev) Non-Commercial License. Commercial use prohibited without authorization.

Citation

@misc{labs2025flux1kontextflowmatching,
      title={FLUX.1 Kontext: Flow Matching for In-Context Image Generation and Editing in Latent Space},
      author={Black Forest Labs et al.},
      year={2025},
      eprint={2506.15742},
      archivePrefix={arXiv},
      primaryClass={cs.GR},
      url={https://arxiv.org/abs/2506.15742},
}