| license: other | |
| library_name: diffusers | |
| pipeline_tag: image-to-image | |
| base_model: | |
| - rimochan/RUM-FLUX.2-klein-4B-preview | |
| base_model_relation: quantized | |
| tags: | |
| - flux | |
| - quantization | |
| - convrot | |
| - int8 | |
| - int4 | |
| # RUM FLUX.2 Klein 4B ConvRot Quantized | |
| Post-training ConvRot quantizations of [RUM-FLUX.2-klein-4B-preview](https://huggingface.co/rimochan/RUM-FLUX.2-klein-4B-preview), based on [FLUX.2-klein-base-4B](https://huggingface.co/black-forest-labs/FLUX.2-klein-base-4B). The repository provides a quality-first INT8 model and a smaller hybrid INT4 model. RUM's dual text-conditioning projections and sensitive input, output, normalization, and modulation layers remain in BF16. | |
| ## Files | |
| - `rum-flux2-int8-convrot-adaround-realcal-hq.safetensors`: INT8 ConvRot + AdaRound with real-activation calibration. | |
| - `rum-flux2-int4-hybrid-convrot-adaround-realcal-hq.safetensors`: hybrid INT4 ConvRot + AdaRound with real-activation calibration; incompatible or quality-sensitive layers remain BF16. | |
| - `standalone_inference.py`: self-contained RUM architecture and quantized runtime entry point for text-to-image and image editing. | |
| Download the FLUX.2 Klein 4B base model and the SDXL/WAI teacher checkpoint separately, then pass their local paths to `standalone_inference.py`. Run `python standalone_inference.py --help` for all options. The recommended defaults are 20 steps for text-to-image and 10 steps for editing. | |
| ## Comparison | |
|  | |