Krea2_FP8

This repository provides an optimized **FP8 (float8_e4m3fn) weight-only quantized version** of the newly released **Krea 2 OSS (Turbo)** transformer. This optimization reduces the model size from the original **24.76 GiB (BF16)** down to **12.01 GiB**, making it highly accessible and runnable on standard consumer hardware (such as 16GB and 24GB GPUs) without sacrificing output quality. - **Original Model Creators**: All credit goes to [KREA.ai](https://www.krea.ai) for the original research, architecture, and weights. - **License**: This model is subject to the **KREA 2 License Agreement**. Please read and comply with the official license terms before using these weights: [KREA 2 Licensing Terms](https://www.krea.ai/krea-2-licensing). - **Purpose**: This repository is a community-contributed utility. It does not claim ownership of the original model or architecture. Its sole purpose is to provide optimized, consumer-hardware-friendly weights for the open-source community. Unlike generic global quantization scripts that aggressively convert every parameter (which often degrades generation details or introduces NaN/promotion calculation errors in neural networks), this model was quantized using a **selective weight-only strategy**: 1. **Targeted Quantization**: Only 2D floating-point weight matrices (`.weight` keys with `ndim >= 2` and element count `> 1024`) were quantized to `torch.float8_e4m3fn`. 2. **Preserved Precision**: - All 1D vectors, biases, and normalization scales are kept in their native high-precision (`float32` / `bfloat16`). - Highly sensitive projection/modulation layers (such as `LastLayer.modulation.lin` vectors) are **completely preserved** in high-precision. This prevents typical mathematical promotion bugs (such as `BFloat16` and `Float8` promotion issues in PyTorch) and retains original output fidelity. 3. **Weight Comparison**: - **Tensors Quantized to FP8**: 266 tensors. - **Tensors Kept in Native Precision**: 166 tensors. - **Size Reduction**: **24.76 GiB ➔ 12.01 GiB** (~51.5% VRAM / disk savings!). Below are official sample outputs from the original Krea 2 OSS Turbo model (generated with the same BF16 weights this FP8 conversion is based on): ComfyUI 0.25.0+ has **built-in Krea2 support**. No custom nodes needed. Drop the workflow JSON into ComfyUI and drag it to the canvas. Place these in your `ComfyUI/models/` folder: Drag [`workflows/Krea 2 simple workflow.json`](workflows/Krea%202%20simple%20workflow.json) onto your ComfyUI canvas. Turbo defaults: **8 steps, CFG 1.0, `er_sde` sampler, `simple` scheduler, 1280×720.** Special thanks to the **KREA.ai** team for releasing Krea 2 to the open-source community. For any commercial licensing inquiries or details about the model, please visit [krea.ai](https://www.krea.ai).