A more structured visualization:
Imposing Consistent Light
In HDR space, illumination has a property that all light transports are independent.
As a result, the blending of appearances of different light sources is equivalent to the appearance with mixed light sources:
Using the above light stage as an example, the two images from the "appearance mixture" and "light source mixture" are consistent (mathematically equivalent in HDR space, ideally).
We imposed such consistency (using MLPs in latent space) when training the relighting models.
As a result, the model is able to produce highly consistent relight - so consistent that different relightings can even be merged as normal maps! Despite the fact that the models are latent diffusion.
From left to right are inputs, model outputs relighting, devided shadow image, and merged normal maps. Note that the model is not trained with any normal map data. This normal estimation comes from the consistency of relighting.
You can reproduce this experiment using this button (it is 4x slower because it relight image 4 times)
Below are bigger images (feel free to try yourself to get more results!)
For reference, geowizard (geowizard is a really great work!):
And, switchlight (switchlight is another great work!):
Model Notes
iclightsd15fc.safetensors - The default relighting model, conditioned on text and foreground. You can use initial latent to influence the relighting.
iclightsd15fcon.safetensors - Same as "iclightsd15fc.safetensors" but trained with offset noise. Note that the default "iclightsd15fc.safetensors" outperform this model slightly in a user study. And this is the reason why the default model is the model without offset noise.
iclightsd15fbc.safetensors - Relighting model conditioned with text, foreground, and background.
Cite
@Misc{iclight,
author = {Lvmin Zhang and Anyi Rao and Maneesh Agrawala},
title = {IC-Light GitHub Page},
year = {2024},
}
Related Work
Also read …
Total Relighting: Learning to Relight Portraits for Background Replacement
Relightful Harmonization: Lighting-aware Portrait Background Replacement
评论