ComfyUI: Flux with LLM, 5x Upscale (Workflow Tutorial)

Seth introduces Flux, a tool with a learning curve that differs from stable diffusion. He explains how Flux deals with pixels rather than resolutions and introduces custom nodes like the Flux Sampler and Flux Resolution Calculator to simplify tasks. Seth demonstrates Flux's ability to generate creative results without adapters or control nets by showing image-to-image transformations based on prompt modifications. He also highlights Flux's capacity for upscaling images and its compatibility with consumer-grade hardware.

The paragraph details the setup process for Flux, including the installation of necessary files and models. It covers the placement of model files in specific folders and the download of upscaler models from Google Drive. Seth advises on the pixel count flexibility of Flux and introduces the Flux Resolution Calculator node, which dynamically adjusts image resolutions. The workflow involves setting and getting nodes to define image dimensions and ratios, ensuring compatibility with Flux.

Seth discusses the use of text conditioning in Flux, where user inputs are structured and expanded by a large language model (LLM). He sets up nodes for text input and connects them to an O Lama Vision node for image reference and image-to-image processes. The paragraph explains the logic behind enabling and disabling LLM chains to optimize the workflow's performance on VRAM. Seth also outlines the use of Boolean logic to control the active mode of connected nodes.

This section delves into the logic behind connecting various nodes for an optimized Flux workflow. Seth uses a Bridge control node to manage inputs and outputs, ensuring that the correct nodes are active based on user selections. He explains the use of Boolean reverse nodes and compares logic nodes to control the workflow direction. The paragraph also covers the creation of custom conditioning to modify prompts without relying on the LLM.

Seth focuses on generating detailed prompts using the text LLM, emphasizing the importance of structured responses for quality outputs. He demonstrates how to use the LLM to create elaborate prompts and then summarizes them for use with the clip L model. The paragraph explains the system instructions for the LLM and the process of testing the logic flow within the workflow.

The paragraph discusses the use of the image LLM for analyzing and describing images in detail. Seth outlines the process of generating a prompt based on the image's composition, style, mood, and colors. He also covers the steps for modifying the prompt using custom conditioning and the importance of maintaining the original context during modifications.

Seth explains the process of modifying image prompts using the LLM and the logic required to enable or disable certain nodes based on user input. He sets up a separate modify image logic to handle user modifications and ensures that the image LLM is only active when needed. The paragraph also covers the use of a switch and compare logic node to control the workflow based on the image LLM's activity.

This section covers the settings and controls for image-to-image processes in Flux. Seth discusses the use of switches and max shift values to control the creative output, allowing for modifications while maintaining consistency with the original image. He also explains the importance of using the correct denoise value and the impact of different max shift values on the final output.

Seth demonstrates how to use inpainting and style transfer techniques in Flux. He shows how to modify prompts and use the inpainting switch to make changes to the image while maintaining consistency. The paragraph also covers the use of different upscalers and the importance of choosing the right model for style transfer and consistency in image details.

The paragraph discusses advanced techniques for upscaling images in Flux, including the use of different upscalers and the importance of maintaining original image consistency. Seth explains the process of calculating new width and height for upscaling and the use of flux samplers to add details. He also covers the use of checkpoints and the importance of selecting the right upscaler for different image types.

Seth describes the final stages of upscaling in Flux, including the use of a 5x upscale without sLing and the reasons behind this choice. He also covers post-processing techniques such as auto-adjustment and levels adjustment to enhance the final image. The paragraph concludes with tips for managing VRAM usage and the potential inclusion of control net support for Flux in future updates.

In the final paragraph, Seth wraps up the tutorial, reflecting on the upscale quality and consistency achieved through the workflow. He mentions the potential for future updates to include support for the Schnell version of Flux and the Guff format. Seth also hints at possible future tutorials and expresses his hope that viewers have gained new insights into using Flux within Comfy UI.