// the find
lllyasviel/IC-Light
More relighting!
IC-Light is a Stanford/lllyasviel research project for relighting a foreground subject with diffusion models, either from a text prompt or a background image. It's aimed at people doing portrait/product photo compositing or ML researchers interested in illumination-consistent generation, not at anyone wanting a packaged app.
The core idea is genuinely interesting: it imposes HDR light-transport consistency in latent space via MLPs, and as a side effect you can extract usable normal maps from relit outputs even though no normal map data was used in training. Ships with two working Gradio demos (text-conditioned and background-conditioned) plus a hosted HuggingFace Space, so you can try it without setting anything up. Comes from a credible author (lllyasviel, of ControlNet fame) with a matching ICLR 2025 paper backing the method, not just a blog post claim.
It's a research drop, not a library: no API, no batch inference script, no packaging beyond the two demo files — you're expected to read gradio_demo.py and hack around it. Background removal depends on BRIA RMBG 1.4, which is non-commercial-only, so anyone wanting to ship this commercially has to swap in BiRefNet themselves and hasn't been given a reference integration. Dependency pinning is loose (raw pip install -r requirements.txt against a CUDA 12.1 torch build), so reproducing the exact demo environment a year+ later is a gamble. No test suite, no CI, no versioned releases — the discussion threads are being used as an ad-hoc changelog instead.