Tagged: FLUX

Recently, ByteDance’s open-source Hyper-sd project officially supports FLUX.1-dev. It currently supports 8-step LoRA and 16-step LoRA, and compared to the default 30-step in FLUX.1-dev, the speed has improved by nearly 4 times! Meanwhile, Hyper-sd’s official team also announced that LoRAs with even fewer generation steps will be released soon, and we’re all looking forward to it!

Introduction

Although quantization is not suitable for conventional UNET models (conv2d), transformer/DiT models like flux seem to be less affected by quantization. This allows them to run in a quantized manner on low-end GPUs. Recent GGUF quantization supports native ComfyUI and FLUX models.

Recently, the original team of SD developed a powerful text-to-image model, Flux.1. Such a strong model cannot afford poor support for Chinese, nor can it lack support for LoRA and ControlNet. Therefore, recently, XLabs-AI has open-sourced the LoRA and ControlNet fine-tuning scripts suitable for Flux.1-dev. In addition, Alibaba Cloud’s Bailing platform launched the optimized Chinese version of Flux on August 8, allowing users to make 1,000 free calls in the first month. Therefore, the following will provide an overview and share insights on these two points.

Everyone is talking about FLUX, and it works great. But for beginners, the first step of downloading and choosing a model can be confusing. There are so many models—how do you know which one to download?

The number of nodes is also overwhelming for beginners.