Flux Kontext [dev]: Custom Controlled Image Size, Complete Walk-through

In this paragraph, the speaker introduces Flux Context, which was recently released just after a live stream where they attempted to emulate the system. The live stream involved experimenting with Flux Context, but the release of the non-API version made much of the previous work obsolete. The speaker highlights the discovery of a potentially useful solution during the live stream, specifically dealing with a feature called flux text image scale, and begins to break down the workings of Flux Context. They explain the components involved, including image stitching and scaling, and touch on the concept of negative prompts and their relevance.

This paragraph shifts focus to how the speaker intends to work with image composites rather than using image stitching. The speaker explains that image stitching can be limiting and unpredictable because it just adds images next to each other. Instead, they use the image composite node, which offers more control and predictability. They walk through the process of loading images, resizing them, and masking out backgrounds to fit one image into another. The speaker also addresses potential issues with dimension mismatchesFlux Context overview and how to overcome them by using a resize node and adjusting the position of the images.

Here, the speaker discusses the challenge of ensuring the proper positioning of the images after they’ve been resized and placed together. They explain how adjusting the image's X and Y coordinates is necessary to make sure key elements, such as a character's face, are not covered up. The speaker explores how to fine-tune the placement of images using the resize node and the limitations of the image's edges. They emphasize the importance of reference latents and how they affect the final composition, ensuring the characters appear as intended.

In this section, the speaker elaborates on the importance of proper prompting within Flux Context. They note that without a prompt, Flux Context doesn't produce useful results, as it will only generate identical outputs. The speaker explains how the model uses reference latents and the dual clip loader in conjunction with prompts to generate meaningful outputs. They demonstrate the importance of guiding the model with specific instructions to get desired results. The paragraph also introduces an example prompt that describes a scene involving a young Asian woman and a man with a top hat, explaining how Flux Context interprets these prompts.

The speaker examines the results after applying a prompt, noting that while the model makes alterations, it may not always stay true to the original character designs. They discuss how the color of the image wasn't specified, leading to some differences in the result. However, the speaker appreciates how the model works with the given instructions. They also touch on the concept of making adjustments, such as converting an image into a vintage photo, and how different techniques in Flux Context allow for continuous refinement of the image, ultimately showcasing the flexibility and potential of the system.

In the final paragraph, the speaker demonstrates how to refine the image further by applying a vintage style effect. They explain how to guide Flux Context by referencing the latest image output, and then instruct the system to transform it into a vintage-style photo from the 1880s. The speaker emphasizes the iterative nature of Flux Context, showcasing how images can be continuously edited and refined. They also touch on an advanced workflow where the system has grouped nodes for ease of editing, highlighting how this approach reduces complexity and improves efficiency. The paragraph concludes with the speaker offering the workflow as a free resource for Patreon supporters.