Untwisting RoPE in ComfyUI: style transfer setup

This tutorial introduces Untwisting RoPE, a frequency-control framework for shared attention in diffusion transformer models, through ComfyUI workflows. The speaker describes it as training-free and says it works with existing models without additional model downloads. Setup involves cloning the custom-node repository into ComfyUI's custom_nodes folder and loading an example workflow.

The Z-Image Turbo example combines a text prompt with a style reference image. The workflow resizes the reference, encodes it into latent data with a VAE, and connects it to the framework's RF inversion components. Examples explore how reference images affect lighting, colors and decoration. The speaker also substitutes native ComfyUI nodes for some image-processing steps.

For Flux-2 Klein's 9-billion-parameter model, the walkthrough covers model-specific block settings and matching the input and style reference dimensions. It distinguishes style transfer from object referencing: the style latent influences appearance, while reproducing a particular monster requires a separate reference through the model's conditioning.

Qwen Image Edit examples apply fantasy and cyberpunk references to an existing scene. The speaker shows how prompt wording changes the result and cautions that outputs vary. The examples distinguish style control from exact character reproduction.