README.md

---
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

![RUM BF16, INT8, INT4 comparison](rum_int4_hybrid_hq_2x4.png)