IC-Light

IC-Light, from Lvmin Zhang (lllyasviel, the ControlNet author; paper 'Imposing Consistent Light' with Maneesh Agrawala, 2024), is a relighting model built on Stable Diffusion v1.5 that manipulates the illumination of an existing image. Give it a foreground subject and a text description of the lighting ('sunlight from the left window', 'warm cinematic') or a background plate, and it re-renders the subject's lighting to match while keeping identity and detail. The conditioning channels were trained specifically for illumination harmonization, which is what separates it from generic img2img: the light behaves like light, with coherent shadows and color temperature.

The v1 models on lllyasviel/ic-light carry a creativeml-openrail-m license, and a v2 series built on FLUX with a 16-channel VAE adds native high resolution and better detail preservation. There are no published benchmark scores, evaluation is demo-driven, and later papers have noted that the training pipeline and dataset were never released, which limits reproducibility. Even so, IC-Light became the default relighting tool in ComfyUI workflows and product-photography pipelines, and the v2 foreground-conditioned model is the one to use for photographic inputs.

No published benchmark scores; evaluated via qualitative relighting demos