Flex.1-alpha 512 Redux Adapter

Overview

This is a Redux adapter trained from scratch on Flex.1-alpha, however, it will also work with FLUX.1-dev. The main advantages of this adapter over the original are:

Better Vision Encoder: The original Redux by BFL uses siglip-so400m-patch14-384. This adapter uses siglip2-so400m-patch16-512, which is newer and has an increased resolution of 512x512. This allows for better image quality and more detail in the generated images.
Apache 2.0 License: Since this vision encoder was trained from scratch on Flex.1-alpha, it can be licensed under the Apache 2.0 license. This means that you can use it for commercial purposes without any restrictions.

The Vision Encoder

Flex Redux uses the latest SigLIP2 vision encoder at a resolution of 512x512. This allows the vision encoder to see more fine details in the images.

ComfyUI

You will need this custom node. https://github.com/ostris/ComfyUI-Advanced-Vision and the siglip2 vision encoder here https://huggingface.co/ostris/ComfyUI-Advanced-Vision/blob/main/clip_vision/siglip2_so400m_patch16_512.safetensors . Put that inside your ComfyUI/models/clip_vision/ folder. If you don't have Flex.1-alpha, you can pick it up here https://huggingface.co/ostris/Flex.1-alpha/blob/main/Flex.1-alpha.safetensors . Put that in ComfyUI/models/checkpoints/. You can also use FLUX.1-dev if you prefer.

Put the flex1_redux_siglip2_512.safetensors file in this repo in ComfyUI/models/style_models/ . There is also a workflow named flex-redux-workflow.json to get you started.

Fine tune it!

This adapter was 100% trained with AI Toolkit on a single 3090 over the course of a few weeks. Which means you can fine tune it, or train your own, on consumer hardware. Check out the example config to get started.