Frequency separation is a native ComfyUI approximation of the Photoshop technique. It splits an image into two layers:
The ComfyUI version comes from Reddit user Powered_JJ's node group. It was posted after Baioni proposed the Photoshop version in answer to a viewer who wanted a whiskey bottle's fine details kept.
After IC-Light relighting (Stable Diffusion 1.5) the low-frequency layer is what should change, because the light should match the new scene. The high-frequency layer should stay the original product's. Splitting both images lets the pipeline keep the new lighting and the real product's detail. This addresses the 'not selling a 90% replica but the actual product' problem in product photography.
The math is tuned for average images and may need adjustment. It also fails on people, because harsh lighting leaves shadow and light information in the detail layers (see ic-light-detail-restore-workflow-limits).
Frequency separation lets the workflow take lighting and detail from different images. The generated image supplies color and lighting (the low-frequency layer). The original supplies fine detail (the high-frequency layer), such as printed text. This is why the product stays the real product after relighting and re-backgrounding.
The mask that selects the subject also decides where this applies. See mask-based-high-frequency-blend. Variations: hand-drawn-detail-preservation-mask and upscaling-placement-in-relight-workflow.