lllyasviel/IC-Light

question about environment light source image L

wang21jun opened this issue · 7 comments

wang21jun commented

Thanks for your great work!
Why the environment light source image L, mentioned as a prompt embedding in your paper, is not included in your infer code?

mhq1065 commented

same question

YeJiangnan1029 commented

Same question. The paper mentions that L is encoded as 3 768-dim tokens, which is the same as the shape of text embedding. Does the light embedding (3 tokens) be removed in the inference stage? Which tokens have been replaced with(or added to) the light embeddings among 77 text tokens in the trainning stage? Or is it posible that make use of the light embeddings(or environment light L) to control the result of inference result? Waiting for your kind reply! :)

zyx1213271098 commented

same question

laolongboy commented

same question

bhosalems commented

Without training code we will not know how the model is trained.

jxl0131 commented

Without training code we will not know how the model is trained.

But if you use L in training, you must use L in inferring